Add from_slow in fast tokenizers build and fixes some bugs

[sgugger]

What does this PR do?

This PR adds an argument to the initialization of the PreTrainedTokenizerFast to force the conversion from a slow tokenizer. This will be useful to help users re-build the tokenizer.json file for some models where we can't update faulty ones right now without breaking backward compatibility (see #9637).

In passing it fixes a few bugs:

• wrong formatting for the documentation
• the fast sentencepiece tokenziers don't have an sp_model attribute so remove the documentation for that
• BarthezTokenizerFast was not registered properly in the autotokenizers, so AutoTokenizer was not finding it

[n1t0]

I say yes!!!

[LysandreJik]

Thank you for applying these changes. They look good to me!