Remove TrainingArguments import from experimental trainers

[albertvillanova]

Remove TrainingArguments import from experimental trainers.

This PR refactors multiple experimental trainer configuration files to remove direct dependencies on transformers.TrainingArguments and instead rely on the project's own _BaseConfig for managing valid dictionary fields. This change improves modularity and reduces external coupling across the configuration classes.

Follow-up to:

• Refactor CLI [14/N] : Remove TrainingArguments import from core trainers #5257

Changes

Dependency and configuration updates:

• Removed the import of TrainingArguments from all experimental config files (BCOConfig, CPOConfig, KTOConfig, OnlineDPOConfig, ORPOConfig) to eliminate unnecessary dependency.
• Changed the assignment of _VALID_DICT_FIELDS in each config class to use _BaseConfig._VALID_DICT_FIELDS instead of TrainingArguments._VALID_DICT_FIELDS, ensuring consistency and reducing external dependency.

---

Note

Low Risk
Low risk refactor that only changes how experimental configs derive _VALID_DICT_FIELDS; runtime impact should be limited to argument/dict field validation if _BaseConfig._VALID_DICT_FIELDS differs from TrainingArguments in a given transformers version.

Overview
Removes the TrainingArguments import from several experimental trainer config modules (BCOConfig, CPOConfig, KTOConfig, OnlineDPOConfig, ORPOConfig).

Each config now sets _VALID_DICT_FIELDS based on _BaseConfig._VALID_DICT_FIELDS (plus model_init_kwargs) rather than TrainingArguments._VALID_DICT_FIELDS, reducing direct coupling to transformers in these experimental configs.

^{Written by Cursor Bugbot for commit c2e93cc. This will update automatically on new commits. Configure here.}

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.