Given a language model in ARPA format, builds a binary representation of the
language model and writes it to disk. Language model binaries are
significantly smaller and faster to load than ARPA files.
Note that if the input/output files have a .gz suffix, they will
be unzipped/zipped as necessary.