Fix stream_redirector_race_condition test
We don't really understand why this test starts failing when we switch
to the new non-blocking worker, as nothing inside the `_ChildWorker`
process should have changed as part of this work.
The failure is due to a failure to drain the wrapped STDOUT/STDERR
streams within the hard-coded one second timeout. Writes occasionally
take nearly 100µs in this test (because the events channel is being
spammed with huge output objects), which is more than enough to make it
hard to flush 10k lines within a second.
Regardless, we'd like to know when this happens, so we've reworked the
drain call to raise a TimeoutError, which we can then catch and reraise,
adding a log line to the relevant prediction before we do so that we can
at least understand what's happening.
We've also just increased how long we're prepared to wait for the drain
-- up to ten seconds -- which we hope will be more than enough for any
non-pathological case.
