Fix DPO completion log-prob loss for unequal target lengths

[TengJiao33]

Summary

• compute DPO preference scores from chosen/rejected completion sequence log-probabilities instead of full-vocab log-prob tensors
• mask prompt and padding tokens so the DPO objective only compares completion tokens
• support chosen and rejected completions with different token lengths by reducing each example to a scalar sequence log-probability
• restore adapter/training state with a finally block after reference-model scoring
• keep public APIs, hparams, and editor call patterns unchanged

Validation

• python -m py_compile easyeditor/models/dpo/dpo_main.py
• temporary regression checks for completion-only masking, shifted token log-prob gathering, and unequal chosen/rejected completion lengths
• smoke edit run confirmed the previous shape mismatch no longer occurs when chosen and rejected targets have different lengths