Don't load entire distribution into memory

[dstufft]

Currently this fetches the distribution from PyPI into a BytesIO object, after doing a requests.get() call (not streaming).

That means that while we're inside of _get_dist, we'll currently be using 2x the file size of the distribution worth of extra RAM, and outside of it we'll be using 1x the file size of extra RAM.

This should probably buffer to a temporary file and use streaming requests so that a large distribution doesn't kill us on memory.

This might just be #5 but I wanted to call it out explicitly since this applies even if we're storing the files somewhere.

[di]

I think we could probably improve this before we get to fixing #5.