mtp: use inp_out_ids for skipping logit computation (#23433)

when doing a follow-up decode for the draft model, we were always doing the logit computation even though it is not required.

src/models/qwen35.cpp

@@ -526,6 +526,7 @@ llama_model_qwen35::graph_mtp::graph_mtp(const llama_model & model, const llm_gr
     res->add_input(std::move(inp));
 
     ggml_tensor * inp_pos     = build_inp_pos();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
     auto * inp_attn           = build_attn_inp_kv();
 
     ggml_tensor * h_norm = build_norm(h_input, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);
@@ -615,6 +616,8 @@ llama_model_qwen35::graph_mtp::graph_mtp(const llama_model & model, const llm_gr
     cb(cur, "h_pre_norm", -1);
     res->t_h_pre_norm = cur;
 
+    cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+
     ggml_tensor * head_norm_w = layer.nextn.shared_head_norm
             ? layer.nextn.shared_head_norm
             : model.output_norm;

src/models/qwen35moe.cpp

@@ -589,8 +589,10 @@ llama_model_qwen35moe::graph_mtp::graph_mtp(const llama_model & model, const llm
     res->add_input(std::move(inp));
 
     ggml_tensor * inp_pos     = build_inp_pos();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
     auto * inp_attn           = build_attn_inp_kv();
 
+
     ggml_tensor * h_norm = build_norm(h_input, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);
     cb(h_norm, "mtp_hnorm", il);
 
@@ -710,6 +712,8 @@ llama_model_qwen35moe::graph_mtp::graph_mtp(const llama_model & model, const llm
     cb(cur, "h_pre_norm", -1);
     res->t_h_pre_norm = cur;
 
+    cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+
     ggml_tensor * head_norm_w = layer.nextn.shared_head_norm
             ? layer.nextn.shared_head_norm
             : model.output_norm;