fix: 学习率调度按 optimizer step 计数并防止 warmup 越界

- total_steps 除以 accumulation_steps，匹配 optimizer.step() 频率
- warmup_steps 用 min 截断，避免 lr_decay_steps 为负

scripts/tools/train.py

@@ -260,13 +260,13 @@ def train(
         },
     )
 
-    total_steps = len(dataset) * n_epoch // (batch_size * nprocs)
+    total_steps = len(dataset) * n_epoch // (batch_size * nprocs) // accumulation_steps
     scheduler_fn = partial(
         create_scheduler,
         **{
             "schedule_type": "cosine",
-            "warmup_steps": warmup_steps,
+            "warmup_steps": min(warmup_steps, total_steps),
-            "lr_decay_steps": total_steps - warmup_steps,
+            "lr_decay_steps": total_steps - min(warmup_steps, total_steps),
         },
     )