Automatically fail over SELECTs to secondaries

Automatically failing over SELECT queries to secondaries would significantly increase read availability when a primary node fails. We can now add secondary nodes to `pg_dist_node`, so the planner could generate task placements for secondaries that are tried after the primary. In addition, we could also load-balance between secondary and primary to reduce load on the primary.

A requirement for read-your-writes consistency is that `synchronous_commit = 'remote_apply'` on all nodes.