minor KTO setting changes + KL batch size

[kawine]

What does this PR do?

• change the default learning rate from 5e-7 to 1e-6; for models 7B and up (which is what most people use w/ or w/o LoRA), i this works better with the default beta=0.1 than an lr=5e-7, which can lead to slow learning
• add explanations in the docs on choosing the learning rate
• (important) change the batch size of the KL data to be the per-step batch size, not the effective batch size; the current implementation leads to worse results when gradient accumulation is used
• allow dropout to be used, like in DPOTrainer

cc @kashif @qgallouedec

[kashif]

in the kto_config.py docstrings can you kindly add:

disable_dropout (`bool`, *optional*, defaults to `True`):
            Whether to disable dropout in the model.

[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.

[kawine]

in the kto_config.py docstrings can you kindly add:

disable_dropout (`bool`, *optional*, defaults to `True`):
            Whether to disable dropout in the model.

done!