🌡️ Fix temperature inconsistency in GRPO trainer

[Aladoro]

What does this PR do?

It fixes the temperature scaling omission when computing the model's and reference model's log probabilities, potentially causing highly biased policy gradients.

This should be consistent with other TRL implementations as PPO and RLOO. Omitting this term has been evidenced to potentially cause detrimental instabilities when fine-tuning LLMs with RL (e.g., see here). Consistently with these results, using my own custom implementation after this fix also seems to work better. I tried doing additional research, but I could not find references that mention purposefully biasing the optimization leads to benefits... am I missing something?

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a GitHub issue? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[qgallouedec]

That's a good point.
That's also what's done in open-instruct: https://github.com/allenai/open-instruct/blob/6d5320539f23a6dd55c892fd35e7e86907569af1/open_instruct/grpo_vllm_thread_ray_gtrl.py#L777C9-L777C37
Ideally, we would like to have some curves to show this gap, so if someone has any, feel free to share.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.