Fixes #209 (again 😓 )

This PR adds a Token class to keep track of whether a token was originally a subword or added token. If it was a subword token, we should not pretokenize it. Previously, we relied on whether a token was in the original vocabulary, but this turned out to not work for unigram tokenizers with a metaspace if the token was also a subword token.

For "ELLE", for example:

1. Check if "ELLE" is in vocab, we pretokenize it to "_ELLE". This is fine
2. Then, when adding it, we need to check whether it is a subword. But "ELLE" is also a subword.
3. So we don't pretokenize it before adding it, causing us to add two tokens with the same surface form.