Because unavailability timer resets every time the controller leader
changes, robustly waiting for the timer to elapse is hard. Instead we
simply wait until the unavailable node appears in the "violations"
status field.

Previously, when the controller leader node was suspended during the
test all status requests would fail with the timed-out error.
This was true for all nodes, not just the suspended one (because we
proxy the status request to the controller leader), so internal retries
in the admin API wrapper didn't help. We increase the timeout and add
504 to retriable status codes so that internal retries can handle this
situation.