llama: avoid copying logits during prompt decode in MTP (#23198)

* llama: avoid copying logits during prompt decode in MTP

* review: update comment

* llama-graph: call set_output for t_h_pre_norm

src/llama-cparams.h

@@ -28,7 +28,8 @@ struct llama_cparams {
     float yarn_beta_slow;
 
     bool embeddings;
-    bool embeddings_pre_norm; // also extract the hidden state before the final output norm
+    bool embeddings_pre_norm;        // also extract the hidden state before the final output norm
+    bool embeddings_pre_norm_masked; // extract for only rows where batch.logits != 0
     bool causal_attn;
     bool offload_kqv;
     bool flash_attn;