Consider computing German fluency without translating to English

[ischender]

As per comment thread here: #69 (comment) :

At the moment, we translate to English before computing fluency. German and English are similar enough that, in the tests I ran, it works better than using not-so-great models specifically for German (eg: something not very grammatical would get a 0.3... fluency score if translated, and a 0.9 using the model).

It could be useful to find/train models that calculate fluency for German without translating, since translation of course loses some of the grammatical fluency bits.