Alternative to #7741

• Move PointerUpdates to its own file, along with tests.
• Introduce opnums for each pointer update, to make sure we remove the exact history that was persisted.
  • Disclaimer: this implementation resets the opnum count once all updates to a single ID have been persisted. This means that it might not behave well if we take two flusher copies of pending changes, and execute them out of order. But this should be already guaranteed by flushing sequence (callers).
• Make drain_persisted to take by value, so it drops directly within the function; at the expense of removing an reinserting updates.