Many of the tests follow a pattern of "wait for N candidates to be
gathered, then (without waiting) assert that gathering is complete". But
this only works if the "gathering complete" signal happens in the same
task as the last candidate being gathered, which isn't an API guarantee.
So the tests will be less fragile if they do the reverse: "wait for
gathering to be complete, then (without waiting) assert that N candidates
were gathered".
Also fixing some somewhat unrelated issues elsewhere. Like a test that
was supposed to be waiting for some period of time and ensuring no
additional candidates were gathered, but wasn't actually waiting at all.
BUG=None
Review-Url: https://codereview.webrtc.org/3018493002
Cr-Commit-Position: refs/heads/master@{#19872}
7f1563facf