Add PSPO trust region method as alternative to clipping in GRPOTrainer

[MCDwyer]

What does this PR do?

Adds PSPO (Probability Smoothing Policy Optimisation) as an alternative trust-region method to GRPOTrainer. PSPO smooths probabilities toward the behaviour policy instead of using ratio clipping.

Paper: https://arxiv.org/abs/2509.21282

Changes:

• Added trust_region_method parameter to GRPOConfig (default: "clip")
• Added smoothing_alpha parameter for PSPO (default: 0.1)
• Implemented PSPO smoothing in GRPOTrainer
• Maintains backward compatibility

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.
@kashif

[qgallouedec]

thanks! can you apply the style (make precommit) and add short section in the paper index documentation page

and also, it would be nice to have a test case for this.

[MCDwyer]

I've applied the style, added the paper documentation in docs/source/gr_pspo.md, and added a test in tests/test_grpo_trainer.py. Please let me know if there is anything else I need to do?

[qgallouedec]

Oh, sorry, maybe I wasn't clear. You need to add a section to this part of the documentation:

https://github.com/huggingface/trl/blob/main/docs/source/paper_index.md

[MCDwyer]

Sorry, I had misunderstood, thank you for clarifying. I've moved the documentation to a section in the paper_index.md