Add tokenizer to Trainer object

[mrm8488]

This way, It will be pushed too on trainer.push_to_hub()

[philschmid]

I kept its purpose separate if you want to "decouple" the training and processing parts.