Allow file-based *2vec training from compressed files

[piskvorky]

All Gensim algorithms allow the use of smart_open to read their input data, meaning the data can be .gz, .bz2, live on s3, etc.

However, the new code path for file-based training of *2vec model from #2127 only accepts .txt files. This is problematic, because the main purpose of this file-based training is to be run on very large datasets (where its superior speed actually matters). Keeping such large text files uncompressed is wasteful and sometimes even impossible.

Task: implement support for reading input from .gz compressed files (at least). bz2 would be nice too, but the "seeking into the middle of a file" by each worker may be problematic for that format, technically speaking.