[WIP] Fix OnlineDPO vLLM server completion handling

[JohnGiorgi]

What does this PR do?

Fixes #5514

This removes an extra flattening step in OnlineDPOTrainer._generate_vllm_server() when using vLLM server mode.

trl/scripts/vllm_serve.py already returns completion_ids as one list[int] per completion. OnlineDPOTrainer was flattening that result a second time, which turned multi-token completions into one-token completions before decode.

This PR removes that second flatten and adds a regression test covering the server return shape.

Validation:

• make precommit
• PYTHONPATH=/tmp/trl-upstream-fix /mnt/home/john/elms-ai/.venv/bin/python -m pytest tests/experimental/test_online_dpo_trainer.py -k preserves_completion_lists

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.
  OnlineDPOTrainer._generate_vllm_server() flattens vllm-serve completion_ids twice #5514
• Did you make sure to update the documentation with your changes?
  No documentation changes are needed for this internal bug fix.
• Did you write any new necessary tests?

AI writing disclosure

• No AI usage: the PR was written entirely by a human.
• AI-assisted: some parts were suggested or improved by AI, but the PR was written and reviewed by a human.
• AI-generated: the PR was mostly or fully generated by an AI tool.

Who can review?

Anyone familiar with OnlineDPOTrainer server-mode generation or the trl vllm-serve response shape.