ClusterTopologyManagerImpl should only use non-blocking RPCs

ClusterTopologyManagerImpl still uses some blocking RPCs, particularly when a node becomes coordinator or after a merge. It should use non-blocking RPCs instead.

It could also use non-blocking RPCs instead of fire-and-forget messages for things like topology updates, which would allow delivering topology updates in the same order on all the nodes instead of having regular nodes make to with missing topology updates (when the coordinator doesn't change).