[MLU] fix: reuse ckpt send/recv buffers to avoid OOM

[uqyxx]

Summary

1. Reuse ckpt send/recv buffers to avoid OOM
2. set_expandable_segments(False) before buffer to avoid memory register error

Testing

• pytest tests/ -v passes
• Manually verified on target hardware (if applicable)

Acceptance Baseline (for new hardware adaptation PRs)

• Ran scripts/baseline_grpo_gsm8k.sh on target hardware (8 devices)
• Training completed all epochs without error
• critic/rewards/mean shows clear upward trend in first 100 steps
• Curve is consistent with NVIDIA reference

SwanLab or training log link:

Reward curve comparison (first 100 steps):

Checklist

• Code follows the project's style and passes pre-commit checks
• Documentation updated (if applicable)
• No secrets or credentials included