Use one full-window running baseline for BCO gradient accumulation

[gss10282025]

What does this PR do?

Collect detached rewards for the whole optimizer window, update RunningMoments once, and let every microbatch read the same updated mean. This keeps the unsplit BCO objective while allowing gradient accumulation.

The pending microbatches are collected in get_batch_samples, their rewards are computed without gradients, and running-statistic writes are suppressed during the loss calls. Evaluation stays read-only. Freezing the old mean and committing after backward would be a different objective, so that is not what this does.

At TRL 1.12.0 the one-vs-four-microbatch parameter-update difference drops from 0.06736848 relative L2 to zero; fixed-baseline and DPO controls agree across layouts; reproduced on a second host. The collection pass restores RNG state and registered model buffers. Full Trainer/GPU execution and multi-rank aggregation still need validation.

Before submitting

• This PR fixes a typo or improves the docs
• Did you read the contributor guideline, Pull Request section?
• Was this discussed/approved via a GitHub issue? Reported in [Bug] BCO uses different running baselines within one gradient-accumulation window #7265; no maintainer approval is implied.
• 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.

Checks on this branch

Added tests/experimental/test_bco_window.py. Its isolated CPU execution passed using the production prefetch/loss methods and RunningMoments: full, equal and uneven windows agree in baseline and gradient; the prepass restores buffers and RNG; evaluation does not update the running state. This isolates the window logic and does not run a complete Trainer job. Ruff, Python syntax and git diff --check pass.

Fixes #7265. The separate removal proposal #7136 is still open; this draft addresses the behavior while BCO remains available.

[qgallouedec]

Closing in favour of #7136, which removes trl.experimental.bco.