Fix dataset preprocessing cache reuse in DPOTrainer and SFTTrainer . Reprocessing the same raw dataset with…

Fix dataset preprocessing cache reuse in DPOTrainer and SFTTrainer . Reprocessing the same raw dataset with…: a task in HF ML Tasksmith (Harbor dataset). Preserve preprocessing behavior for both plain-text and conversational datasets, including prompt/completion boundaries, preference outputs, EOS…