This PR is in preparation for data quarantining. The DQ PR will move the
tables affected by this PR entirely into memory, and crash recovery will
be driven by re-processing consensus commits.

However, when we deploy DQ to a validator for the first time, we will
restart in a state where uncertified consensus commits have already been
marked as processed, so they will not be re-processed on startup.

This means that upon starting up for the first time, some data will be
on disk instead of in memory. There are two possible solutions to this:

- All reads fall back to the database. This is a lot of ugly code, and
is slow.
- All data in the database is read into memory at startup. This is fast
and simple.

This PR bounds the amount of data we will have to read at startup in
order to make the second option feasible.