Allow passing kwargs through to TFBertTokenizer

[Rocketknight1]

There are some kwargs like preserve_unused_tokens in the underlying TF tokenizer layers that might be useful to expose to users. This PR exposes them by passing through any unrecognized kwargs in the model __init__ to the TF tokenizer layer.

Fixes #23798

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[Rocketknight1]

Ping @amyeroberts for core maintainer review now that the extra functionality is working fine (see issue #23798)

[amyeroberts]

Thanks for adding!

In general, we definitely don't want to add more **kwargs arguments. As this is a wrapper for a TF implemented tokenizer, I think it's OK. Only niggle is it might result in some conflicting behaviour e.g. does FastBertTokenizer take any arguments which are equivalent to e.g. truncation ?

[Rocketknight1]

I was a bit wary about the kwargs thing too - FastBertTokenizer and BertTokenizerLayer actually have wildly different arguments, so depending on which one you're using the kwargs you need will be totally different. Still, I think for an advanced use case it's fine - we're just trying to enable some power user behaviours without forcing them to edit the library source, and I'd prefer something general like this over specifically exposing the options I think people need (because I didn't even realize in advance that the preserve_unused arg would be valuable!)

Anyway, merging!