Our tests running via RunWithLocks strictly depend on no
watcher jobs interfering. Therefore they pause the watcher;
unfortunately, there still is a race: the watcher only checks
the pause status upon its invocation, but submits jobs later
in its run time. Therefore not only pause it (doesn't hurt),
but also add a filter to reject all its jobs, and then wait
for all running jobs to terminate.

Signed-off-by: Klaus Aehlig <aehlig@google.com>
Reviewed-by: Petr Pudlak <pudlak@google.com>