[draft] Use vLLM in LogCompletionsCallback

[tchang1997]

What does this PR do?

Implements changes to completion-logging callback suggested in #2786.

Normally, LogCompletionsCallback calls a helper function called _generate_completions. However, this uses the default model.generate, which can be slower than vLLM. This PR:

• adds a helper called _generate_completions_vllm with equivalent functionality
• adds a switch to on_step_end in LogCompletionsCallback to choose whether to use vLLM or not based on args.use_vllm from the Trainer
• adds a utility function to [partially] convert the GenerationConfig normally passed to __init__ in the callback to SamplingParams compatible with vLLM.

I have not fully written tests yet (hence the draft), or fully checked compatibility with non-GRPOTrainer trainers, but this yields identical logging as the original LogCompletionsCallback in my own training scripts. Posting PR draft here for early feedback.

Fixes # (issue)

#2786

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.
• [in progress] 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.