Make vector loading more flexible.

[anna-hope]

• Add support for passing a max_vectors argument, in situations where the entire set of pre-trained vectors can't fit in memory (most pre-trained vector sets are sorted in the descending order of word frequency, so if you pass, say, max_vectors=100_000, you will get pre-trained for 100_000 most common words)
• Add ability to read pre-trained vectors from gzipped files (without tar)
• Read pre-trained vectors by iterating over the vectors file object line by line, without reading the whole file into memory first
• Only load the file in binary mode to make parsing logic more compact
• Make vector loading more memory-efficient (based on Fix vector loading memory requirements #298)
• Minor refactoring to improve style

[mttk]

@notnami could you fix the issues Travis has with code style so this can be merged? (there's just a few indentations and unused imports)

[anna-hope]

I fixed the issues with style, but now the Python 3.5 build has failed for an unclear reason. I believe that same build had passed before. Someone with write access to the repo could restart it, or I could close this PR and re-submit.

[mttk]

@notnami I restarted it, probably just timed out. Will merge when it passes.

[mttk]

@notnami thanks a lot!