[Feature] Support Distributed LogProb for GRPO Training

[duanjunwen]

📌 Checklist before creating the PR

• I have created an issue for this PR for traceability
• The title follows the standard format: [doc/gemini/tensor/...]: A concise description
• I have added relevant tags if possible for us to better distinguish different PRs
• I have installed pre-commit: pip install pre-commit && pre-commit install

🚨 Issue number

Link this PR to your issue with words like fixed to automatically close the linked issue upon merge

e.g. fixed #1234, closed #1234, resolved #1234

📝 What does this PR do?

Summarize your work here.
if you have any plots/diagrams/screenshots/tables, please attach them here.

💥 Checklist before requesting a review

• I have linked my PR to an issue (instruction)
• My issue clearly describes the problem/feature/proposal, with diagrams/charts/table/code if possible
• I have performed a self-review of my code
• I have added thorough tests.
• I have added docstrings for all the functions/methods I implemented

⭐️ Do you enjoy contributing to Colossal-AI?

• 🌝 Yes, I do.
• 🌚 No, I don't.

Tell us more if you don't enjoy contributing to Colossal-AI.

[TongLi3701]

Thanks Junwen, I left some comments.

[TongLi3701]

Please also compare the peak memory when using these two different methods.

[duanjunwen]

Please also compare the peak memory when using these two different methods.

We compared the case where the parallel output is True or False under strategy tp2zero1.
When parallel output is False, the peak mem is 93000+ MB (may suffer OOM when other users grab resources).
When parallel output is True, the peak mem is around 82000 MB.

[TongLi3701]

Thanks Junwen, I left some comments. Please address the comments and click on resolve conversation then request review again.

Thanks.

[duanjunwen]

Resolve Conflict.