models : support qwen3.5 series (#19468)

* support qwen3.5 series

* remove deepstack for now, and some code clean

* code clean

* add FULL_ATTENTION_INTERVAL metadata

* code clean

* reorder v heads for linear attention to avoid expensive interleaved repeat

src/llama-model.h

@@ -118,6 +118,7 @@ enum llm_type {
     LLM_TYPE_21B_A3B, // Ernie MoE small
     LLM_TYPE_30B_A3B,
     LLM_TYPE_31B_A3_5B,
+    LLM_TYPE_35B_A3B, // Qwen3.5
     LLM_TYPE_48B_A3B, // Kimi Linear
     LLM_TYPE_80B_A3B, // Qwen3 Next
     LLM_TYPE_100B_A6B,
@@ -322,6 +323,9 @@ struct llama_layer {
     // qwen3next
     struct ggml_tensor * ssm_beta_alpha = nullptr;
 
+    // qwen3.5
+    struct ggml_tensor * ssm_alpha = nullptr;
+
     // rwkv
     struct ggml_tensor * time_mix_w1         = nullptr;
     struct ggml_tensor * time_mix_w2         = nullptr;