WIP: Reduce flapping in replicaiton tests by waiting for replication to finish

This uses a local IRPC message to determine when the replication pull queue is empty, and so avoid issues with a replication run continuing after replication has been disabled locally.

To make this work reliably in the test system an issue, unexpected in the original design, in the messaging subsystem needs to be addressed. Because the client messaging context cleanup (close()/unlink()) happens on TALLOC_FREE(), a second use of the per-pid cached FD on the server-side can fail, so we need to catch that, re-try and re-open the new socket.