Scaling out single node Citus to a multi-node cluster might lead to higher planning times for OLTP workloads

With #3388, we enabled plan caching when a the shards reside on the node itself that plans the distributed query. The meaning of that PR is Citus caches the plans for the shard queries as well. With a single node Citus, all shards would be avaliable locally, meaning that prepared statements would be as performant as Postgres. However, when the cluster is scaled out, this optimization would be lost.

So, a user starts with a high performant single-shard OLTP workload, and scaling out would add significant increase in the latencies (as far as I remember ~%30).

Note that due to network round-trip, the execution would be definitely more expensive. This issue is about planning times, as the execution time increases seems inevitable.