fix: don't re-flatten vLLM server completion_ids in Online DPO

[vineethsaivs]

What does this PR do?

OnlineDPOTrainer._generate_vllm_server re-flattened the vLLM client's completion_ids:

completion_ids = [[comp_id] for prompt_completions in completion_ids for comp_id in prompt_completions]

VLLMClient.generate(...)["completion_ids"] already returns a list[list[int]] with one token-id list per completion (the same shape the colocate path produces and that GRPO uses). The comprehension iterated over every completion and every token, so a completion like [101, 102, 103] became three single-token "completions" [101], [102], [103]. Downstream that makes completion_mask [1, 0, 0, ...] for every row and throws off the per-process row count. The colocate path (_generate_vllm_colocate) and GRPOTrainer never apply this extra flatten, so server-mode Online DPO was the only inconsistent path. This removes the re-flatten so completion_ids is passed through unchanged, and adds a CPU regression test.

This supersedes the abandoned draft #5516 (a WIP with the same one-line removal, untouched since April); I commented on the issue first to coordinate.

Fixes #5514

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. See OnlineDPOTrainer._generate_vllm_server() flattens vllm-serve completion_ids twice #5514.
• Did you make sure to update the documentation with your changes?
• 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 in the community is free to review the PR once the tests have passed.

---

Note

Low Risk
Small behavioral fix in experimental Online DPO vLLM server path only; aligns with colocate behavior and is covered by a targeted unit test.

Overview
Fixes vLLM server-mode Online DPO generation where _generate_vllm_server incorrectly re-flattened VLLMClient.generate’s completion_ids. The client already returns list[list[int]] (one full token sequence per completion, same as colocate/GRPO); the removed comprehension treated each token as its own completion, breaking completion_mask and per-process row counts.

_generate_vllm_server now passes completion_ids through unchanged after the main-process generate call. A CPU regression test (test_generate_vllm_server_preserves_completion_token_lists) stubs gather/broadcast and the vLLM client to assert multi-token completions stay intact.

^{Reviewed by Cursor Bugbot for commit 6a9f32a. Bugbot is set up for automated code reviews on this repo. Configure here.}