Trainer trains the Model in-place

[n1t0]

Current state

When training a Tokenizer, its Model gets replaced after training since the Trainer generates a new Model. This has several limitations:

• When the Model being replaced has been customized (dropout, unk_token, ...), we lose all of this when we replace it (cf BPE dropout not working as expected #201)
• In Python, if we keep a reference to the model added to the Tokenizer, this reference does not point to the actual model used by the Tokenizer after training.

Goal

Change the Trainer to actually train to Model in-place.