Imporve NMT quality by adding 10-100 new tokens with SentencePiece

[johnml1135]

It is assumed that at least for some languages, adding 10-100 new compiled tokens (depending potentially on the number of instances, the size of corpus, etc.) may end up improving the overall training quality. This issue is to implement it after the research is complete (see sillsdev/silnlp#196).