As discovered in bgo#665728, all our prepare() (and unprepare()) methods are
currently vulnerable to being run multiple times concurrently from a single
thread. Add a _pending_prepare flag to all of them to prevent this. It
doesn't need to be locked, since it should only ever be accessed from a
single thread (since only a single thread can get through the lock{}
recursive mutex at once).

Helps: bgo#665728