updated BERT processing to current interface

[piotrpiekos]

Adapts old code to new interface. For now it's just fine-tuning bert of chosen glue dataset without evaluating on test set

I didn't use SubwordTextEncoder because it gives different token_ids on words that consist of more than one token (e.g. "neglecting").
I didn't upgrade tfds_stream to data_stream, because it might be necessary in BERT models to be able to do processing after batching, so this is to be revisited after implementing more BERT family models.

[afrozenator]

Hi @piotrekp1 - Thanks for adding this! I'll accept it right now without changes, but in a future PR could you move it to trax/data instead and add a simple unit test there? (That package should have examples tests with testdata that will make this easy).

[afrozenator]

@henrykmichalewski as FYI

[piotrpiekos]

Hi @afrozenator. I'll do that, thank you!