Check tokenize params on DPOTrainer

[pablovicente]

Given that we tokenize the dataset on the DPOTrainer.__init__ regardless of having a custom data_collator, we need to ensure that a tokenizer is passed and warm about the need to set max_length, max_prompt_length and max_target_length . Hence, moving those checks outside of the data_collator check.

On older implementations these parameters were only needed if the user did not provide a data_collator but they are now required for all DPO runs.

@kashif

[HuggingFaceDocBuilderDev]

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.