Turns out for long-running upgrades we were running them all in
parallel, which thrashes the disk pretty hard.  This adds a simple mutex
lock around each upgrade, so at least the computer is usable while it's
going on.  A more robust solution would be to have a single-thread
thread pool where we enqueue upgrades, but that's too much change this
late in the release cycle.

Also it turns out that the nullifying of the internaldate_time_t column
before we repopulate it was very costly, and unnecessary.  So, this also
should speed things up for upgrading users.

Closes: bgo #724475