adding Huggingface compatible tokenizers

[justin-barton]

Adding AptTokenizer() and EsmTokenizer() based on the Rust PreTrainedTokenizerFast class and original EsmTokenizer class, respectively.

Some benefits of this change are:

• multiprocessing for distributed tokenization
• batch_decode methods
• attention_masks and special_tokens_masks
• simplifies dependencies (doesn't require rust_trie)
• enables easy distribution on Huggingface Hub later

As described in #47

[jamaliki]

Hi @justin-barton ,

Thank you so much for your contribution! We discussed this change and we are going to put this on the back burner for now, since we get some flexibility benefits from having the tokenizer completely defined within the repo in this early stage of training. However, for later steps in the project, such as when we release models to the HF hub, we will probably end up using this PR to have a fully HF-compatible tokenizer.

Best,
Kiarash.