fix(npu): synchronize stream after backward for dp replicate training

[danieldale2026]

Summary

Fixes #59.

This PR fixes loss=nan in NPU training when dp_replicate_size > 1 by synchronizing the current NPU stream immediately after backward.

Details

• Adds torch_npu import in mova/engine/trainer/accelerate/accelerate_trainer.py
• Calls torch_npu.npu.current_stream().synchronize() right after self.accelerator.backward(loss)
• Keeps the change scoped to NPU stream ordering after backward
• Does not modify other training logic

Test

• python3 -m compileall mova/engine/trainer/accelerate/accelerate_trainer.py
• git diff --check