* m_aDeadProcessors was useless; for one, removeEventsForProcessor could never
  have run (and set m_aDeadProcessors) between execute's reading from aEvents
  and checking m_aDeadProcessors (because of the use of aMutex in both
  functions), and if that were addressed, there would always be a race that
  execute would run a processor that has just been removed.  Clients have to be
  aware that a call to removeEventsForProcessor is just an optimization hint,
  but does not guarantee that the given processor is not called from the execute
  thread after removeEventsForProcessor returns.

* The sequence

    aGuard.clear(); aPendingActions.reset(); aPanedingActions.wait();

  could cause calls to aPendingActions.set() to get lost, both for signalling
  new content in the queue and for signalling termination.

Change-Id: I43293e3d5090c2d46db8bc8ed6fb9c71e049d55c