GitLab

Tokenization behave the same as original XLM proprocessing for most languages except zh, ja and th; Change API to allow specifying language in `tokenize`