Cleaner Tokenization Handling in tidybert::tidy_bert_output()

[jonthegeek]

The basic_usage.Rmd vignette still has to do a manual tokenize step. That shouldn’t be necessary.

Right now we only auto-tokenize with {luz} fit/predict (and the callback telling it to do that). We need a clean way to tokenize when we use pretrained berts more directly, like we do here.

If that worked, we could create an (untokenized) dataset, then use it in the model, at which point it would be updated to match the model (or we call a helper or whatever). And then tidy_bert_output() could accept a dataset_bert_pretrained for its second argument.