Support offset mapping alignment for fast tokenizers

[adrianeboyd]

Switch to offset mapping-based alignment for fast tokenizers. With this change, slow vs. fast tokenizers will not give identical results with spacy-transformers.

Additional modifications:

• Update package setup for cython
• Update CI for compiled package

[adrianeboyd]

Concerns:

• ~~use_fast should be saved correctly in the tokenizer settings (as a separate PR, #339)~~
• ~~naming of _align and new methods~~
• whether we (initially) want an automatic backoff to the old alignment if the new alignment fails in some cases
• further speed improvements
  • in particular I suspect this should be improved: https://github.com/explosion/spacy-transformers/pull/338/files#diff-f767b104e773bafeb21b7870c06043e8cebea5c4fce193b171b737192d589f62R90

[adrianeboyd]

Yes, this would need to be v1.2 because it will change the behavior of trained models.

[adrianeboyd]

In terms of model performance and speed this implementation appears to be on par with the existing spacy-alignments alignment. In general I think it should able be to be little bit faster, my first guess (as mentioned above) is now here in the renamed version:

https://github.com/explosion/spacy-transformers/pull/338/files#diff-b326f7a5839c4589ffcac094a207251f14e6235d150b35a7ae07459776bcc19aR104

I'm trying to find a reasonable test case for where the actual alignments would make a difference, but most things I've tried so far were a wash. The main known cases are with roberta/gpt2 with the bytes_to_unicode mapping. At least users looking at the alignments would be less concerned about it?