Optimization: Qwen3 next autoregressive pass (ggml-org#17996)

* It's Qwen3 Next, the lean mean token generation machine!

* Apply patches from thread

* Remove recurrent version, only keep chunked and autoregressive

* Remove unnecessary conts and asserts

* Remove more extra conts and asserts

* Cleanup masking

Author: pwilkin

src/models/models.h

@@ -441,13 +441,14 @@ struct llm_build_qwen3next : public llm_graph_context_mamba {
                 ggml_tensor * cur,
                 ggml_tensor * causal_mask,
                 ggml_tensor * identity,
+                ggml_tensor * diag_mask,
                         int   il);
 
     ggml_tensor * build_layer_ffn(
                 ggml_tensor * cur,
                         int   il);
 
-    ggml_tensor * build_delta_net_recurrent(
+    ggml_tensor * build_delta_net_chunking(
                 ggml_tensor * q,
                 ggml_tensor * k,
                 ggml_tensor * v,
@@ -456,18 +457,17 @@ struct llm_build_qwen3next : public llm_graph_context_mamba {
                 ggml_tensor * state,
                 ggml_tensor * causal_mask,
                 ggml_tensor * identity,
+                ggml_tensor * diag_mask,
                         int   il);
 
-    ggml_tensor * build_delta_net_chunking(
+    ggml_tensor * build_delta_net_autoregressive(
                 ggml_tensor * q,
                 ggml_tensor * k,
                 ggml_tensor * v,
                 ggml_tensor * g,
                 ggml_tensor * beta,
                 ggml_tensor * state,
-                ggml_tensor * causal_mask,
-                ggml_tensor * identity,
-                        int   il);
+                int           il);
 
     ggml_tensor * build_norm_gated(
                 ggml_tensor * input,