Ensure that the stream feeder doesn't hang in write...

...if the restoring data engine thread winds up operations.  By closing
the engine side of the pipe unconditionally when exiting the thread,
the unanticipated-failure path is now guaranteed (instead of blocking
forever in write() to a pipe that isn't being read!).

In addition, wire agent-timeout handling into the various stream
data-handling operations (preflight, backup, restore).  This were
not sufficiently robust and were in some situations leaving the
backup/restore mechanisms in a livelock state.

Finally, plug a longstanding problem in which we'd have orphaned
timeout messages coming in and producing a certain amount of "wtf?"
logging and wasted CPU.  No longer!

Bug 28457158

Change-Id: I597c76c3eada378ffeb20870253847594f73e089