Save tokenizer as part of model

The tokenizer for a given model is deterministic (it only depends on the vocab file + whether it's cased). Producing the tokenizer takes 100x as long as loading a pre-processed tokenizer (about 4 s vs 40 ms for bert_base_uncased).

Save the tokenizer as part of the download process. If a model has a vocab but not a tokenizer, save a tokenizer once and then use it going forward (for backward compatibility with things that are already downloaded).