[prf/dec][doc] Update javadoc to reflect unified batched prefill-decode plan

Author: orionpapadakis

src/main/java/org/beehive/gpullama3/tornadovm/layers/type/fp16/decode/LlamaFP16FFNLayersDecode.java

@@ -9,7 +9,8 @@
 import uk.ac.manchester.tornado.api.enums.DataTransferMode;
 
 /**
- * Decode-path FFN layers for the Phase 4 unified plan.
+ * Decode FFN layers of the unified batched prefill-decode plan
+ * ({@link org.beehive.gpullama3.tornadovm.TornadoVMMasterPlanWithBatchPrefillDecode}).
  *
  * <p>Overrides data-transfer declarations so that all cross-graph boundaries use
  * the explicit-source form of {@code consumeFromDevice}.  The no-arg form (used by
@@ -20,11 +21,6 @@
  * never propagated — causing either a null-pointer crash or a silent re-upload
  * from host (zeros), corrupting the hidden state and KV cache.</p>
  *
- * <p>Two boundaries are fixed here:</p>
- * <ul>
- *   <li>{@code wrapX}: via {@link #predecessorGraphName} hook in the base class.</li>
- *   <li>All other consumed objects: via the {@link #configureLayerDataTransfers} override.</li>
- * </ul>
  */
 public class LlamaFP16FFNLayersDecode extends LlamaFP16FFNLayers {
     public LlamaFP16FFNLayersDecode(String taskGraph, LlamaState state,

src/main/java/org/beehive/gpullama3/tornadovm/layers/type/fp16/decode/LogitsFP16LayerDecode.java

@@ -8,7 +8,8 @@
 import uk.ac.manchester.tornado.api.TaskGraph;
 
 /**
- * Logits layer for the unified prefill-decode plan (Phase 4).
+ * Logits layer of the unified batched prefill-decode plan
+ *  * ({@link org.beehive.gpullama3.tornadovm.TornadoVMMasterPlanWithBatchPrefillDecode}).
  *
  * <p>Extends {@link LogitsFP16Layer} with KV-cache pass-through so the device
  * pointers for {@code wrapKeyCache} and {@code wrapValueCache} survive the

src/main/java/org/beehive/gpullama3/tornadovm/layers/type/fp16/prefill/LlamaFP16LayersBatchPrefill.java

@@ -16,7 +16,8 @@
 import java.util.stream.IntStream;
 
 /**
- * Builds per-layer batch prefill TaskGraphs for Phase 4 GPU batched prefill.
+ * Prefill FFN layers with batching for the unified batched prefill-decode plan
+ * ({@link org.beehive.gpullama3.tornadovm.TornadoVMMasterPlanWithBatchPrefillDecode}).
  *
  * <p>One {@link ImmutableTaskGraph} per transformer layer, each processing
  * {@code batchSize} tokens simultaneously via {@link TransformerBatchPrefillKernels}.</p>