Don't crash when backup timeout races with agent completion

There's a narrow window of time in which an agent reporting that its
operation has completed races with timeouts such that we wind up
handling the completion callback just after certain fundamental state
has been reset.  Detect this race and proceed gracefully instead of
crashing.

Bug 19498669

Change-Id: I5a475527db1a55a8e567366ddfb10112e427682e