1. We shouldn't run the callback inside mask_, because that prevents it
   from receiving StackOverflow, amongst other things.  I think this was
   a  n attempt to prevent the delivery of ThreadKilled inside the
   callback, but to do that you would need uninterruptibleMask_
2. killThread doesn't wait for the threads to die, so use the async
   package and cancel/wait instead.
3. If the killThread happens during a callback, we don't want to discard
   the exception, because that will leave the thread running.
I've added a new test for the problem of a callback hanging.