feat: add EAGLE3 speculative decoding support

EAGLE3 is an encoder-decoder based speculative decoding method:
- Extracts features from target model at specific layers
- Uses feature fusion layer to compress target features
- Generates draft tokens with single-layer decoder
- Maps draft vocabulary to target vocabulary via d2t tensor

Key changes:
- Add LLM_ARCH_EAGLE3 architecture
- Add EAGLE3 encoder/decoder graph (src/models/eagle3.cpp)
- Add feature extraction from target model layers
- Add g_embeddings handling for decoder input
- Add GGML_TENSOR_FLAG_SYNC for GPU synchronization
- Add --eagle3 flag for speculative-simple example
- Add EAGLE3 model conversion in convert_hf_to_gguf.py

include/llama.h

@@ -363,6 +363,13 @@ extern "C" {
         bool kv_unified;  // use a unified buffer across the input sequences when computing the attention
                           // try to disable when n_seq_max > 1 for improved performance when the sequences do not share a large prefix
                           // ref: https://github.com/ggml-org/llama.cpp/pull/14363
+        
+        // EAGLE3 extraction configuration
+        // When eagle3_model is set, layer extraction is automatically enabled
+        const struct llama_model * eagle3_model; // EAGLE3 model to read extract_layers configuration from
+                                                 // If non-NULL, enables automatic feature extraction
+        const struct llama_model * target_model; // reference to target model
+                                                 // only used to share embedding layer with eagle3 model
     };
 
     // model quantization parameters
@@ -846,6 +853,23 @@ extern "C" {
                     llama_seq_id   dest_seq_id,
            llama_state_seq_flags   flags);
 
+    //
+    // EAGLE3 draft model support
+    //
+
+    // Get pointer to target model features extracted for EAGLE3 encoder
+    // Returns NULL if no features are available
+    // Format: [3*n_embd, n_tokens] - use model.hparams.n_embd and batch.n_tokens for dimensions
+    LLAMA_API const float * llama_get_eagle3_target_features(struct llama_context * ctx);
+    
+    // Set g_embeddings from EAGLE3 encoder output for decoder input
+    // g_embd: pointer to encoder output embeddings
+    LLAMA_API void llama_set_eagle3_g_embeddings(
+            struct llama_context * ctx,
+                   const float * g_embd,
+                       int32_t   n_embd,
+                       int32_t   n_tokens);
+
     //
     // Decoding
     //