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

[onyxmaster]

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.