Notable changes:

* Relation forks, Free Space Map rewrite, and Visibility Map. This
  conflicted quite heavily with the file mirroring stuff in GPDB. The file
  mirroing is kept unchanged as much as possible, and the extra FSM and VM
  forks are not mirrored.

* Introduce SEMI and ANTI joins, and some other subselect-related planner
  changes. We had already backported these changes earlier, so we had most
  of these changes already.

* Implement WITH, including WITH RECURSIVE. We already had this, and we had
  already refactored the old GPDB implementation to match the upstream one,
  so we had most of this already.

* Change parsing of INTERVAL literals to be spec-compliant. This fixes
  github issues #1187 and #3023.

* Make LC_COLLATE and LC_CTYPE per-database settings. We had some code
  related to file replication that checked the cluster's LC_COLLATE/LC_CTYPE
  settings, but that no longer makes sense because they're now per-database.