Align KTO with DPO: Fix KTO compute_metrics type hint (EvalLoopOutput → EvalPrediction)

[qgallouedec]

Part of the effort to align KTOTrainer (experimental) with DPOTrainer. #4786

KTOTrainer.__init__ typed compute_metrics as Callable[[EvalLoopOutput], dict], which was both wrong (the Trainer passes an EvalPrediction to compute_metrics, not an EvalLoopOutput) and inconsistent with the trainer's own docstring, which already documents it as EvalPrediction. Switched the annotation and the import to EvalPrediction, matching DPOTrainer.

No behavior change (type-annotation only).

---

Note

Low Risk
Type-annotation-only change in experimental KTO trainer; no logic, data, or API behavior is modified.

Overview
Aligns experimental KTOTrainer with DPOTrainer by correcting the compute_metrics typing only.

The import and __init__ annotation now use EvalPrediction instead of EvalLoopOutput, matching what Hugging Face Trainer actually passes into compute_metrics and what the class docstring already described. No runtime or training behavior changes.

^{Reviewed by Cursor Bugbot for commit 35a7703. Bugbot is set up for automated code reviews on this repo. Configure here.}

[bot-ci-comment]

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.