fix(train): align DDP no_sync with gradient accumulation

[YuCeong-May]

Summary

Type of change

• Bug fix
• Documentation
• Example or demo
• Runtime or deployment
• Benchmark or evaluation
• Model/training change

Validation

• python -m compileall funasr examples tests
• Docs or links checked
• Runtime/deployment command tested

User impact

Notes for reviewers

[LauraGPT]

Verified against current main and with a two-rank CPU/Gloo DDP reproduction.

The old batch_idx % accum_grad condition synchronizes the first microbatch and suppresses synchronization on the optimizer-step microbatch, leaving rank-local accumulated gradients (4.0 vs 7.0 in the reproduction). The proposed (batch_idx + 1) % accum_grad condition keeps early microbatches inside no_sync() and synchronizes the final microbatch, producing identical accumulated gradients on both ranks (5.5 vs 5.5).

Additional checks:

• synthetic merge with current main is clean and changes only this condition
• python -m py_compile funasr/train_utils/trainer_ds.py
• git diff --check

Approved.