fix: F4-1/2/3 — three crash-on-run bugs

F4-1 (dashboard.py): _init_tokenizer() -> _ensure_tokenizer()
F4-2 (train_grpo.py): GRPOConfig — remove non-existent fields
(learning_rate/reward_fn_name/num_generations)
F4-3 (train_grpo.py): GRPOTrainer(config=...) -> grpo_config=...;
extract learning_rate into separate AdamW optimizer;
pass reward_fn name string to trainer (resolved at call time)
8/8 tests passed.

inference/dashboard.py

@@ -77,7 +77,7 @@ class InferenceEngine:
         self.model.eval()
         self.inference_config = InferenceConfig()
         self.kv_cache = None
-        self._init_tokenizer()
+        self._ensure_tokenizer()
         
     def _load_model(self, model_path: str) -> tuple:
         """Load model from path."""

train/train_grpo.py

@@ -126,18 +126,22 @@ def main():
     # 加载模型
     model, config, device = load_fusion_model(args.model_path, args.device)
 
-    # GRPO 配置
+    # GRPO config
     grpo_config = GRPOConfig(
-        learning_rate=args.learning_rate,
         kl_coef=args.kl_coef,
-        reward_fn_name=args.reward_fn,
-        num_generations=args.num_generations,
     )
 
-    # 创建 GRPOTrainer
+    # Optimizer (separate from GRPO config)
+    optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
+
+    # Reward function (string name, resolved by GRPOTrainer.compute_reward)
+    reward_fn = args.reward_fn
+
+    # Create GRPOTrainer
     trainer = GRPOTrainer(
         model=model,
-        config=grpo_config,
+        grpo_config=grpo_config,
+        reward_fn=reward_fn,
     )
 
     # 加载数据