Reshard with logical replication, part I

[levkk]

Description

Implement logical replication and COPY fan-out between N origin shards and M destination shards. This reshards an existing cluster, one table at a time and streams table updates to the right shard(s) by parsing the replication stream.

Implementation

1. Using binary COPY, stream tables to destination cluster while sharding between nodes.
2. Using logical replication, stream changes to each table and convert logical repl. messages into upserts using prepared statements.

Upserts allow us to rewind the stream back and replay without duplicates. All tables have to have a primary key, which will be used as the replica identity. DO NOT use the whole row as a replica identity, that doesn't work for upserts.

Missing pieces

1. Foreign keys not supported (easy fix, see below)
2. Schemas not automatically synchronized, requires manual pg_dump/pg_restore (see below)
3. Updates and deletes (easy fix, see below)

To solve the FK problem, we need to sync the schema ourselves without the constraints (or additional indexes). Once the replication is caught up, re-create the constraints & add back the indexes.

Updates and deletes just need us to hook into the query parser a bit to get shard on the where clause. This is still a PoC, this will be done in the next PR.

Misc

1. Connections from connection pools that have replication_mode enabled will be closed instead of re-used. This ensures that connections left in partial state after logical replication are not re-used by other clients.
2. Add Cluster::execute method to execute a query on all primaries in the cluster. Just a helper, useful for testing,
3. Add Pool::standalone to create a connection untracked by the pool. Can be used for whatever reason, incl. replication.