Add support for periodic reference model synchronization (TR-DPO style) to `KTOConfig` and `KTOTrainer`, matching the existing DPO implementation.

**KTOConfig changes**

Add three new fields with the following defaults:

| Field | Type | Default |
|---|---|---|
| `sync_ref_model` | `bool` | `False` |
| `ref_model_mixup_alpha` | `float` | `0.6` |
| `ref_model_sync_steps` | `int` | `512` |

When `sync_ref_model=True`, the reference model is periodically updated: `π_ref = α · π_θ + (1−α) · π_ref_prev`, where α is `ref_model_mixup_alpha` and the update happens every `ref_model_sync_steps` training steps.

**KTOTrainer changes**

With `sync_ref_model=True`, KTO training must update the reference model parameters using the interpolation and training-step interval specified above. The reference parameters must remain unchanged between synchronization steps.

Trainer construction must reject synchronization with a PEFT active model by raising `NotImplementedError`, and reject synchronization with `precompute_ref_log_probs=True` by raising `ValueError`.

When `sync_ref_model=False` (the default), behaviour is unchanged, including support for PEFT and precomputed reference log probabilities.

Work in `/workspace`. Submit your fix in the existing Python source files under `trl`. Preserve the other public behavior. The environment is offline; dependencies are preinstalled. Grading runs the relevant repository tests in a fresh environment, using your submitted source files.