refactor: 训练循环改为两重迭代并统一参数命名

- 训练循环从三重(epoch→batched→batch)改为二重(epoch→batch)
- batch_size → batch_per_device, accumulation_steps → grad_accum_steps
- scheduler 移入 step block 对齐 optimizer 更新步
- GradientClippingCallback 改用 on_step_begin 避免零梯度裁剪
- 移除 _train_impl 误导性的 -> Checkpoint 标注
- total_steps 修除为向下取整并精简为一行
- warmup_steps 改为 warmup_ratio (默认0.05)

astrai/trainer/train_context.py

@@ -70,7 +70,7 @@ class TrainContextBuilder:
         context.scheduler = self.config.scheduler_fn(context.optimizer)
 
         cfg = self.config
-        sampler_offset = context.iteration * cfg.batch_size
+        sampler_offset = context.iteration * cfg.batch_per_device
         sampler = ResumableDistributedSampler(
             data_source=cfg.dataset,
             start_epoch=context.epoch,
@@ -79,7 +79,7 @@ class TrainContextBuilder:
         )
         context.dataloader = DataLoader(
             cfg.dataset,
-            batch_size=cfg.batch_size,
+            batch_size=cfg.batch_per_device,
             sampler=sampler,
             num_workers=cfg.num_workers,
             pin_memory=cfg.pin_memory,