fix(trainer): correct LR schedule, final checkpoint and STG under DDP

[blue (Blueyyds)]

Three defects that only surface on multi-GPU runs or with checkpoint retention enabled:

1. LR scheduler advanced once per process. Accelerate's default step_scheduler_with_optimizer=True applies its distributed batch-size adjustment on top of the trainer's own explicit scheduler.step() per optimizer step, so an 8-GPU run consumed the schedule 8x too fast and hit the final LR after 1/8 of training. The trainer owns the stepping, so opt out.

2. Final checkpoint could delete itself. When the last optimizer step also lands on a save interval, the same path was written twice and appended to the retention list twice; with keep_last_n=1 the pruning pass then removed the file it had just written. Skip the redundant save when the interval checkpoint already exists, with a wait_for_everyone() so all ranks agree on the state.

3. STG validation crashed under DDP. transformer.num_blocks is not reachable through the DistributedDataParallel wrapper, so enabling stg_scale raised AttributeError during validation. Read the attribute off .module when wrapped.