Langid: combine with discriminating word sets

[unhammer]

C.f. #207 we would get even better langid if we combine with hand-collected sets of discriminating words (e.g. "ikke" is never nynorsk, "ikkje" is always nynorsk), see also https://github.com/google-research-datasets/TF-IDF-IIF-top100-wordlists