I noted in D9453 that approach was simpler, but it turns out there is
still a race in that approach. Running each set of tests where nodes are connected
in a hub-and-spoke is more robust, since it's possible for interconnected nodes
to sync blocks at inconvenient times and trigger a reorg. This wasn't caught earlier
because the test happens to complete relatively quickly and the nodes tend to not
sync blocks on their own before the entire test completes.

This approach has slightly more overhead (1 more node), but otherwise functionally
the same as before.