restore Qwen3ForCausalLM architecture (HF supports it natively)

config.json

@@ -1,6 +1,6 @@
 {
     "architectures": [
-        "Qwen2ForCausalLM"
+        "Qwen3ForCausalLM"
     ],
     "attention_bias": false,
     "attention_dropout": 0.0,
@@ -13,7 +13,7 @@
     "initializer_range": 0.02,
     "intermediate_size": 12288,
     "max_position_embeddings": 40960,
-    "model_type": "qwen2",
+    "model_type": "qwen3",
     "num_attention_heads": 32,
     "num_hidden_layers": 36,
     "num_key_value_heads": 8,