Tests pass on pypy

[msabramo]

This is #425 applied to master.

Make test_socket_manager tests pass on PyPy

by applying patch from @pjenvey on IRC:
http://paste.ofcode.org/a588X46F8gNdFDSHDtK5NP

Cc: @mcdonc, @gcarothers, @mnaberez , @pjenvey

[msabramo]

Passing Travis CI build: https://travis-ci.org/Supervisor/supervisor/builds/23315595

Note that this passes on pypy now!. Thanks, @pjenvey!

It's an allowed failure on Python 3.4 because Travis CI doesn't support Python 3.4 yet. But when they add it (hopefully soon), we will see whether it works or not and then we can remove it from the list of allowed failures.

[mnaberez]

Why does it call gc.collect three times in a row? Did it just start working when more calls were added or is there a rationale behind it? I don't see any reason not to merge this, but it would probably be a good idea to put a comment in there explaining it.

[msabramo]

Hopefully @pjenvey can respond with the rationale on the 3x gc.collect. I more or less blindly applied his patch. 😄

[mcdonc]

Thanks!

[pjenvey]

The 3 collect's are along the lines of what the stdlib's test.support.gc_collect does, its docstring has a bit of an explanation:

https://bitbucket.org/mirror/cpython/src/c8ce5bca0fcda4307f7ac5d69103ce128a562705/Lib/test/support/__init__.py?at=default#cl-1359

Basically a 'full collection' with CPython's ref counting GC cleans up most objects immediately, whereas you don't get such a guarantee w/ a non ref counting GC. 3 calls pretty much does it for PyPy. The JVM's GC is even more fussy as you can see gc_collect makes an additional sleep on it