Separate translations into different files

and maybe make them real JSON while we're at it and sort them to be able to keep track of what's present or not in different languages by simple diff-ing. Probably should accept existing MRs to avoid conflicts and extra work.