Standardise API for WinRateCallback and LogCompletionsCallback

[lewtun]

What does this PR do?

This PR unifies the API for the WinRate and LogCompletion callbacks (previously, the latter required the prompts to be passed manually, sampling was tied to the training config etc). The new way to use them is as follows:

judge = PairRMJudge()
generation_config = GenerationConfig(max_new_tokens=256)
win_rate_callback = WinRateCallback(
    judge=judge, trainer=trainer, generation_config=generation_config, num_prompts=32
)
log_callback = LogCompletionsCallback(trainer=trainer, generation_config=generation_config, num_prompts=8)
trainer.add_callback(win_rate_callback)
trainer.add_callback(log_callback)

Doing so allowed me to also fix a bug we had in LogCompletionCallback where ZeRO-3 would hang during generation (the trick is to access and unwrap the model from the trainer with unload_model_for_generation())

Scripts tested

• DPO (DDP & Z3)
• GKD
• XPO
• Online DPO

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.

[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.

[lewtun]

I switched to this dataset since the original doesn't have an eval split

[lewtun]

I unified the import path for the callbacks too

[qgallouedec]

LGTM, it looks way better now!