BERT implementation

[jbarrow]

Added and documented the BERT implementation. It currently uses the HF tokenizer, because that handles a lot of nice things (like padding, masking, and token_type_ids).

It's tested against the HuggingFace implementation at: batch sizes >= 1, different models (bert-base-uncased, bert-base-cased, bert-large-uncased, bert-large-cased), etc.

[awni]

This is really nicely done!! I love it. I left some mostly minor comments.

Please address and then we can merge it into the examples.

Also could you add a requirements.txt for dependenies?

[jbarrow]

Outside of anything that requires a change to MLX core, I think I've made the changes.

I'll put off removing the MultiHeadAttention class until I've made that change, it's been pulled into MLX, and there's been a released version. 😄

[jbarrow]

Submitted and tested a PR for the bias change in mlx. I can submit a new PR here once it's merged/released.

[awni]

Ok, just two more comments, then we can merge this and follow up after the change to core multiheadedattention.

[awni]

Brilliant, thank you for putting this together!