[simplify] Set can_return_loss=True globally, instead of via the data collator

[tomaarsen]

Hello!

Pull Request overview

• Set can_return_loss=True globally instead of via the data collator
• Also override the typings for args and data_collator for easier intellisense etc.

Details

The global can_return_loss should always be used unless the data has a return_loss sample, which we specifically forbid via validate_column_names. Sentence Transformer models can always return a loss during training, so this helps simplify our code somewhat.

• Tom Aarsen