fix: add pull to solve the problem that fsdp2 traninng in shared will timeout

[gygdh-001]

During multi-node multi-GPU FSDP training, checkpoint saving and resume frequently hit timeout errors.

The root cause is that accelerate_trainer.py uses FullStateDictConfig / FullOptimStateDictConfig for FSDP,
which gathers the full state dict from all ranks onto rank 0 during every save_state / load_state call.
Transferring the full parameter set across nodes over the network causes timeouts,
and rank 0 is also at risk of OOM.

This PR applies the following changes:

1. Replaces FullStateDictConfig / FullOptimStateDictConfig with
   ShardedStateDictConfig / ShardedOptimStateDictConfig for FSDP,
   so each rank only holds its own shard without cross-node full-state gather.

2. Under FSDP, save_checkpoint now saves per-rank files:
   optimizer_{rank}.bin and scheduler_{rank}.bin,
   bypassing accelerator.save_state() which would trigger full-state gather.

3. Under FSDP, _resume_checkpoint now loads per-rank optimizer and scheduler
   state dicts independently, bypassing accelerator.load_state() and its full-state gather path.

Validation:

Verified on a multi-node (2 nodes × 8 Ascend 910B) FSDP training setup:

• Checkpoint save time reduced from 400s+ to under 300s, no more timeouts.
• Resume success rate 100%, each rank correctly restores optimizer/scheduler state.
• Resume time reduced from 160s to 40s.
• Loss continuity verified — the loss curve after resume is seamless with the original training run.

[gygdh-001]

@yhzx233