Turns out that when PM is in use, some D3cold resume paths
could have one thread, holding the network lock, deadlock
in flush_scheduled_work() since an event task is waiting
for that same lock.  Fix is to call that later, when the
lock isn't held.