Improve boosting for exact matches with underscores

[kueda]

https://api.inaturalist.org/v1/search?q=helen_y does not include https://www.inaturalist.org/people/helen_y despite the exact match but https://api.inaturalist.org/v1/search?q=helen_y&sources=users does. Most of the results start with helen* so it's possible the underscore is being treated as a token separator, so while I'm guessing this is a problem with our very imperfect "exact" matching, it might also be work looking into how the tokenizer is working.