feature: Configurable num logprobs in vLLM generation

[LeonEricsson]

Summary

• Adds a logprobs parameter across the vLLM generation stack (client, server, and generation orchestrator), allowing callers to request the top-N log-probabilities per generated token rather than only the sampled token's log-probability.
• VLLMGeneration::generate returns a new logprob_token_ids field alongside logprobs, both with shape (num_sequences, seq_len, num_logprobs), giving downstream consumers access to the full top-k distribution.

Motivation

This change is a prerequisite for an upcoming PR I’m working on.

More broadly, it is a strict improvement over the original implementation. Previously, logprobs was hardcoded to 0, which discarded the richer top-k information that vLLM already computes internally. I was lazy when I implemented this the first time around because we only needed sampled logps. By exposing the logprobs parameter, we incur no additional overhead while unlocking access to information that was previously inaccessible.

Minimal upstream effect, GRPOTrainer and RLOOTrainer use the default logprobs=0

todo

• run modified slow test, requires multiple accelerators

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?
• 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.