Unify Android JNI to single IRunner, wire prefill to runner (pytorch#17756)

Replace the dual-runner pattern (runner_ + multi_modal_runner_) with a
single IRunner* that holds either TextLLMRunner or MultimodalRunner,
leveraging MultimodalRunner's new IRunner inheritance from pytorch#17741.

Each prefill method (text, images, audio) now immediately calls
IRunner::prefill(vector<MultimodalInput>) instead of buffering inputs
for later consumption by generate(). A needs_bos_ flag tracks whether
the next prefill should apply BOS tokens — MultimodalRunner also guards
this via pos_==0 internally, but TextLLMRunner trusts the caller.

generate(), stop(), load(), and reset() no longer branch on
model_type_category_; all dispatch through the unified runner_.

Rename all JNI native methods from append* to prefill* to match the
existing Java public API naming.

Author: kirklandsign

extension/android/executorch_android/src/main/java/org/pytorch/executorch/extension/llm/LlmModule.java

@@ -373,19 +373,18 @@ public int generate(
   }
 
   /**
-   * Prefill a multimodal Module with the given images input.
+   * Prefill the KV cache with the given image input.
    *
    * @param image Input image as a byte array
    * @param width Input image width
    * @param height Input image height
    * @param channels Input image number of channels
-   * @return 0, as the updated starting position in KV cache of the input in the LLM is no longer
-   *     exposed to user.
+   * @return 0 on success
    * @throws RuntimeException if the prefill failed
    */
   @Experimental
   public long prefillImages(int[] image, int width, int height, int channels) {
-    int nativeResult = appendImagesInput(image, width, height, channels);
+    int nativeResult = prefillImagesInput(image, width, height, channels);
     if (nativeResult != 0) {
       throw new RuntimeException("Prefill failed with error code: " + nativeResult);
     }
@@ -434,7 +433,7 @@ public void prefillImages(ByteBuffer image, int width, int height, int channels)
     }
     // slice() so that getDirectBufferAddress on the native side returns a pointer
     // starting at the current position, not the base address.
-    int nativeResult = appendImagesInputBuffer(image.slice(), width, height, channels);
+    int nativeResult = prefillImagesInputBuffer(image.slice(), width, height, channels);
     if (nativeResult != 0) {
       throw new RuntimeException("Prefill failed with error code: " + nativeResult);
     }
@@ -500,128 +499,125 @@ public void prefillNormalizedImage(ByteBuffer image, int width, int height, int
     }
     // slice() so that getDirectBufferAddress on the native side returns a pointer
     // starting at the current position, not the base address.
-    int nativeResult = appendNormalizedImagesInputBuffer(image.slice(), width, height, channels);
+    int nativeResult = prefillNormalizedImagesInputBuffer(image.slice(), width, height, channels);
     if (nativeResult != 0) {
       throw new RuntimeException("Prefill failed with error code: " + nativeResult);
     }
   }
 
-  private native int appendImagesInput(int[] image, int width, int height, int channels);
+  private native int prefillImagesInput(int[] image, int width, int height, int channels);
 
-  private native int appendImagesInputBuffer(ByteBuffer image, int width, int height, int channels);
+  private native int prefillImagesInputBuffer(
+      ByteBuffer image, int width, int height, int channels);
 
-  private native int appendNormalizedImagesInputBuffer(
+  private native int prefillNormalizedImagesInputBuffer(
       ByteBuffer image, int width, int height, int channels);
 
   /**
-   * Prefill a multimodal Module with the given images input.
+   * Prefill the KV cache with the given normalized image input.
    *
    * @param image Input normalized image as a float array
    * @param width Input image width
    * @param height Input image height
    * @param channels Input image number of channels
-   * @return 0, as the updated starting position in KV cache of the input in the LLM is no longer
-   *     exposed to user.
+   * @return 0 on success
    * @throws RuntimeException if the prefill failed
    */
   @Experimental
   public long prefillImages(float[] image, int width, int height, int channels) {
-    int nativeResult = appendNormalizedImagesInput(image, width, height, channels);
+    int nativeResult = prefillNormalizedImagesInput(image, width, height, channels);
     if (nativeResult != 0) {
       throw new RuntimeException("Prefill failed with error code: " + nativeResult);
     }
     return 0;
   }
 
-  private native int appendNormalizedImagesInput(
+  private native int prefillNormalizedImagesInput(
       float[] image, int width, int height, int channels);
 
   /**
-   * Prefill a multimodal Module with the given audio input.
+   * Prefill the KV cache with the given preprocessed audio input.
    *
    * @param audio Input preprocessed audio as a byte array
    * @param batch_size Input batch size
    * @param n_bins Input number of bins
    * @param n_frames Input number of frames
-   * @return 0, as the updated starting position in KV cache of the input in the LLM is no longer
-   *     exposed to user.
+   * @return 0 on success
    * @throws RuntimeException if the prefill failed
    */
   @Experimental
   public long prefillAudio(byte[] audio, int batch_size, int n_bins, int n_frames) {
-    int nativeResult = appendAudioInput(audio, batch_size, n_bins, n_frames);
+    int nativeResult = prefillAudioInput(audio, batch_size, n_bins, n_frames);
     if (nativeResult != 0) {
       throw new RuntimeException("Prefill failed with error code: " + nativeResult);
     }
     return 0;
   }
 
-  private native int appendAudioInput(byte[] audio, int batch_size, int n_bins, int n_frames);
+  private native int prefillAudioInput(byte[] audio, int batch_size, int n_bins, int n_frames);
 
   /**
-   * Prefill a multimodal Module with the given audio input.
+   * Prefill the KV cache with the given preprocessed audio input.
    *
    * @param audio Input preprocessed audio as a float array
    * @param batch_size Input batch size
    * @param n_bins Input number of bins
    * @param n_frames Input number of frames
-   * @return 0, as the updated starting position in KV cache of the input in the LLM is no longer
-   *     exposed to user.
+   * @return 0 on success
    * @throws RuntimeException if the prefill failed
    */
   @Experimental
   public long prefillAudio(float[] audio, int batch_size, int n_bins, int n_frames) {
-    int nativeResult = appendAudioInputFloat(audio, batch_size, n_bins, n_frames);
+    int nativeResult = prefillAudioInputFloat(audio, batch_size, n_bins, n_frames);
     if (nativeResult != 0) {
       throw new RuntimeException("Prefill failed with error code: " + nativeResult);
     }
     return 0;
   }
 
-  private native int appendAudioInputFloat(float[] audio, int batch_size, int n_bins, int n_frames);
+  private native int prefillAudioInputFloat(
+      float[] audio, int batch_size, int n_bins, int n_frames);
 
   /**
-   * Prefill a multimodal Module with the given raw audio input.
+   * Prefill the KV cache with the given raw audio input.
    *
    * @param audio Input raw audio as a byte array
    * @param batch_size Input batch size
    * @param n_channels Input number of channels
    * @param n_samples Input number of samples
-   * @return 0, as the updated starting position in KV cache of the input in the LLM is no longer
-   *     exposed to user.
+   * @return 0 on success
    * @throws RuntimeException if the prefill failed
    */
   @Experimental
   public long prefillRawAudio(byte[] audio, int batch_size, int n_channels, int n_samples) {
-    int nativeResult = appendRawAudioInput(audio, batch_size, n_channels, n_samples);
+    int nativeResult = prefillRawAudioInput(audio, batch_size, n_channels, n_samples);
     if (nativeResult != 0) {
       throw new RuntimeException("Prefill failed with error code: " + nativeResult);
     }
     return 0;
   }
 
-  private native int appendRawAudioInput(
+  private native int prefillRawAudioInput(
       byte[] audio, int batch_size, int n_channels, int n_samples);
 
   /**
-   * Prefill a multimodal Module with the given text input.
+   * Prefill the KV cache with the given text prompt.
    *
    * @param prompt The text prompt to prefill.
-   * @return 0, as the updated starting position in KV cache of the input in the LLM is no longer
-   *     exposed to user.
+   * @return 0 on success
    * @throws RuntimeException if the prefill failed
    */
   @Experimental
   public long prefillPrompt(String prompt) {
-    int nativeResult = appendTextInput(prompt);
+    int nativeResult = prefillTextInput(prompt);
     if (nativeResult != 0) {
       throw new RuntimeException("Prefill failed with error code: " + nativeResult);
     }
     return 0;
   }
 
   // returns status
-  private native int appendTextInput(String prompt);
+  private native int prefillTextInput(String prompt);
 
   /**
    * Reset the context of the LLM. This will clear the KV cache and reset the state of the LLM.