[chore] If Transformers 4.46.0, use processing_class instead of tokenizer when saving

[tomaarsen]

Follow-up of #3026

Hello!

Pull Request overview

• If Transformers 4.46.0, use processing_class instead of tokenizer when saving

Details

If we use tokenizer with Transformers v4.46.0+, then we'll get this twice every time that a model is checkpointed

Trainer.tokenizer is now deprecated. You should use Trainer.processing_class instead

• Tom Aarsen