Auto-load processing_class in PRMTrainer when omitted

[DaoyuanLi2816]

What does this PR do?

Auto-loads processing_class from the model when it is omitted, matching SFT, DPO, and Reward trainers. This applies with either the default or a custom data collator and for both raw and pre-tokenized datasets.

PRMConfig now exposes trust_remote_code for tokenizer loading.

Tests

• processing-class autoload coverage for all three dataset/collator paths (3 passed)
• Ruff check and format check on the changed files

AI writing disclosure

• AI-assisted: some parts were suggested or improved by AI, but the PR was written and reviewed by a human.

[qgallouedec]

thanks. I think to better align with the other trainer, a better path would be to automatically load the processor in all cases. see how it's done in SFT, DPO etc.

[DaoyuanLi2816]

Updated this to auto-load the tokenizer whenever processing_class is omitted, including custom collators and pre-tokenized datasets. Added coverage for all three paths.

[DaoyuanLi2816]

@qgallouedec this now loads the processor automatically in all cases, as you suggested, rather than only for the collator path — it covers custom collators and pre-tokenized datasets too, with coverage for all three. Ready for another look whenever you have time.

[qgallouedec]

Closing in favour of #7133, which removes trl.experimental.prm. The autoload is right and matches what SFT, DPO and Reward do, there is just no trainer left to carry it. thks for the work on it.