We were sending one per fork, but a little bit of refactoring allows us
to send just one request with forknum == InvalidForkNumber.  This not only
reduces pressure on the shared-memory request queue, but saves repeated
traversals of the checkpointer's hash table.