SentencePiece is not working properly for the Sinhala Language due to Zero Width Joiner is getting removed

[sarubi]

As per the reported issue, Empty tokens in output vocabulary #276, Zero Width Joiner character has been replaced by whitespace. Due to that languages that require Zero Width Joiner getting altered and resulted wong output. This issue will be there for languages like Sinhala, Devanagari, Kannada and Malayalam [1].

Ideally, we shouldn't replace Zero Width Joiner (200D) with whitespace or empty since it indicates to join two chars without zero width(no whitespace). Also, It requires to present in order to decode the segmentation to raw test successfully. We should keep these special characters as it is.

[1] https://en.wikipedia.org/wiki/Zero-width_joiner