handlers: fix RequestHandler queue flush on stop

[yungchin]

This resolves #309 by removing the weakref solution introduced in #258,
which would potentially throw a ReferenceError before the worker was
quite done emptying the request queue.

[yungchin]

Not to be too much of a codecov cheerleader, but their patch-coverage feature just allowed me to spot a bug that was hidden from the test suite: https://codecov.io/github/Parsely/pykafka/compare/f0f12a3...2d4a35a showed that there's only partial branch coverage for the while not shared.ending: line, which was due to me introducing a bug in the finaliser (and exceptions in finalisers are suppressed, so it went otherwise unnoticed).

@stevepeak

[stevepeak]

👍 Thank you @yungchin Very happy to provide this solution for you!

[emmettbutler]

@yungchin this looks good. I'm pretty happy with any fix that removes a weakref, as those generally make the code harder to reason about. I spent a bit trying to come up with a better name for shared that's more descriptive of what it's doing, but I can't.

[yungchin]

@emmett9001 thanks as ever for the speediness!! I actually think that shared is the worst name possible, but still couldn't think of a better one. It's like they say http://martinfowler.com/bliki/TwoHardThings.html