clean up some runtime potential bugs

Summary:
1. Zero cache on allocation (line 65-66): std::fill on cache_data_ and update_data_ after allocator_.allocate() —eliminates uninitialized memory garbage that varies across devices.
  2. Zero cache on reset (line 191): std::fill on cache_data_ in reset() — ensures stale KV cache from a previous prompt is fully cleared, not just the position counters.
  3. Zero padding in last prefill chunk (line 618-621): When batch_len < input_len, fill the tail of the input buffer with zeros — prevents stale tokens from a previous chunk leaking through the embedding layer. sa_runner.cpp 
  4. Call runner.reset() before each prompt in the multi-prompt loop, stdin prompt loop, and stdin tokens loop —ensures the KV cache, masks, and input_pos_ are fully reset between prompts

Differential Revision: D104615993

Author: billmguo

examples/models/llama/runner/static_attention_io_manager.h

@@ -64,6 +64,8 @@ class StaticKVCache {
     update_data_ = allocator_.allocate(update_data_size_);
     ET_CHECK(cache_data_ != nullptr);
     ET_CHECK(update_data_ != nullptr);
+    std::fill(cache_data_, cache_data_ + cache_data_size_, T(0));
+    std::fill(update_data_, update_data_ + update_data_size_, T(0));
     init_ptrs();
   }
 
@@ -186,6 +188,7 @@ class StaticKVCache {
    */
   void reset() {
     std::fill(cache_pos_.begin(), cache_pos_.end(), 0);
+    std::fill(cache_data_, cache_data_ + cache_data_size_, T(0));
   }
 
  private:
@@ -613,6 +616,10 @@ class StaticAttentionIOManager {
         return config_.generate_full_logits ? input_len - 1 : 0;
       }
       std::copy(&tokens[i], &tokens[i + batch_len], input_buffer.begin());
+      if (batch_len < input_len) {
+        std::fill(
+            input_buffer.begin() + batch_len, input_buffer.end(), TokenT(0));
+      }
       if (!config_.generate_full_logits && config_.last_valid_token_pos_index) {
         last_valid_token_pos_ = batch_len - 1;
         set_input(