Use load based strategy for selecting readonly connections

[lutter]

When using read replicas, we right now select the database for a connection in a round robin fashion. But since the primary will always have to do more work (indexing and serving subscriptions), the utilization of the read replica(s) will have a ceiling well short of 100%.

Instead of round-robin, we should select read replicas by using the pool with the smallest connection wait time as that is a decent proxy for database load.