Fix regression in task stealing for already released keys

[fjetter]

This fixes a regression introduced in #4107

I believe this scenario can only be encountered in some race condition scenarios which is why I had a hard time constructing a test case which is less close to the implementation.

[gforsyth]

This fix looks correct to me -- I hadn't hit this particular edge case while working on #4107 but the key in self.tasks check is prevalent throughout. The Travis failure is spurious -- if someone can restart the windows 3.6 build, I'm curious to see if that is a legitimate failure or just a bit of flakiness.

[jrbourbeau]

if someone can restart the windows 3.6 build, I'm curious to see if that is a legitimate failure or just a bit of flakiness

I pushed an empty commit so we can still see the initial CI failure (hope that's okay @fjetter)

[gforsyth]

Great. Looks like the windows failure was a fluke -- once this passes on Travis should be good to go in.

[fjetter]

Builds still fail and run into connection timeout errors. I'm wondering if this connects to #4176 :/

[mrocklin]

Thank you for identifying and fixing this @fjetter

Builds still fail and run into connection timeout errors. I'm wondering if this connects to #4176 :/

In situations like these I recommend submitting a fresh PR to test CI and see if the issue repeats there. If so, we should raise an issue, but can go ahead and merge this in.

[jrbourbeau]

There's already an issue for test_broken_worker_during_computation (xref #4173), so I'll go ahead and merge this PR in