Align KTO with DPO: Precompute reference log probs at init.

This PR refactors the logic for precomputing reference log probabilities in the KTOTrainer class. The main improvement is the consolidation of duplicated code for precomputing reference log probabilities for both training and evaluation datasets into a single reusable method, which simplifies the codebase and reduces redundancy.

Part of:

• KTO refactoring #4786

Motivation

KTOTrainer precomputed reference log probs lazily inside get_train_dataloader and get_eval_dataloader overrides, guarded by _precomputed_train_ref_log_probs / _precomputed_eval_ref_log_probs flags.

DPOTrainer already uses the correct pattern: precompute in __init__ via a dedicated _precompute_ref_logps helper, so no dataloader overrides are needed at all.

This PR aligns KTOTrainer with that approach:

• Remove get_train_dataloader and get_eval_dataloader overrides
• Remove _precomputed_train_ref_log_probs and _precomputed_eval_ref_log_probs flags
• Add _precompute_reference_log_probs(dataset, batch_size, desc) -> Dataset helper
• Call it from __init__ (after super().__init__(), once self.accelerator is available) for both train and eval datasets

Changes

Refactoring and code simplification:

• Removed the duplicated logic from get_train_dataloader and get_eval_dataloader by introducing a unified _precompute_reference_log_probs method that handles reference log probability computation for any given dataset.
• The new _precompute_reference_log_probs method is now called during initialization if precompute_ref_log_probs is enabled, updating both train_dataset and eval_dataset as needed.

Cleanup:

• Removed unused flags and state variables (_precomputed_train_ref_log_probs, _precomputed_eval_ref_log_probs) and related comments, as precomputation is now handled directly and only once during initialization.