The XLM-Roberta-Base model is not 100% compatible with HF Tokenizers implementation

Please check the repro at https://github.com/onyxmaster/BlingFireIssue/.
It is able to correctly tokenize the string `¯` to tokens `6`, `166757`, but fails on the `¯¯`, tokenizing them as `6`, `166757`, `166757` instead of `6`, `166757`, `6`, `166757`.