This is a proposed add-on / follow-up to SERVER-126806.

With SERVER-126806, the catalog's timestamp-aware cache is populated at startup (at recoveryTimestamp) then kept up to date with DDL operations. However to serve reads at t < recoveryTimestamp we'd need to still need to have a fallback path which does a catalog full scan by NSS/UUID then populates the HistoricalCatalogIdTracker.

However all data needed to populate the timestamp-aware cache up from the beginning of the _mdb_catalog WiredTiger snapshot history exists on disk. We can read history a WiredTiger  dump_version cursor and put it in the cache by committing it at a past commit_timestamp.

This has the following advantages:

• Avoid catalog full scans to make catalog performance predictable.
• Remove the complexity of the HistoricalCatalogIdTracker.
• Unify all catalog accesses into a single path.

The main roadblock is that WiredTiger's dump_version API is documented as a debug feature. We should discuss if that can be made acceptable for us, or find an alternative.