Fix fasttext model loading from gzip files

[mpenkov]

Due to a bug in numpy (numpy/numpy#13470), we are unable to load models from gzip files. @menshikh-iv came up with a workaround: bypass the numpy.fromfile function, do the file I/O ourselves, and pass an iterator to numpy.fromiter.

• Fix Using the load_facebook_model method produces ValueError on array reshaping #2473

[menshikh-iv]

It's a bit better to have something like

if fname.endswith("gz"):
    # this will be here until https://github.com/numpy/numpy/issues/13470 will be fixed
    <new_custom_code_same_as_in_pr>
else:
    <old_code>

This allows to

• don't reduce performance (time & RAM) in the case when file is uncompressed (because it will work in the exact same way)
• avoid the current numpy issue with gz FP

[piskvorky]

The workaround in this PR looks clean and simple, I like it. What are its performance implications?

IIRC, loading FT models is already slower than Facebook's tool, so we have to be careful not to introduce extra penalties. Or be upfront about the trade-offs, perhaps even suggesting to the user to decompress their model instead (if the implications are too dire).

[mpenkov]

It takes 2min 23s to load the uncompressed model (commoncrawl, 4GB). It takes 4min30s to load the uncompressed model (7GB).