Alternative approach to fixing Poller races/hangs

In case it might be helpful, here is my current best functioning (read: least hang-y :zany_face:) attempt at a fix for `Poller`. The tl;dr is that with these changes, the IJulia PR gets through all tests (with only a couple failures, including a couple timeout's aka kernel/poller hangs), but only with verbose/debug logging. I believe that confirms that the issue isn't with the IJulia eventloop flow, and suggests that there is still a race(?) somewhere in the `Poller` implementation. (Replaces #266 )