The background migration MigrateMergeRequestDiffCommitUsers processes
batches of merge_request_diff_commits rows in batches. These migrations
have been running for several weeks now.

Unfortunately, on GitLab.com we are running into problems with the
remaining 750 or so jobs. Some jobs take a few minutes to complete,
while others can take over an hour to complete. The scheduling interval
was originally just a few minutes. This lead to many jobs being thrown
away, as they could not be picked up fast enough when there is a long
running job. Rescheduling jobs with a greater interval hasn't proven
successful.

In this commit we add a new background migration to clean up the pending
work. This background migration picks up a single job (in reverse
order), processes it, then schedules the next job. This approach
prevents jobs from conflicting with each other if they take too long.

In addition, we change the existing migration to not process jobs that
have already been processed. This way existing jobs that have yet to be
picked up won't lead to duplicate work.

Both the old and new migration _can_ process work in parallel, as they
are different background migration classes. This shouldn't pose any
problems. For the last few weeks we've been manually running jobs in a
Rails console in parallel to the Sidekiq jobs, without any problems
surfacing.

See https://gitlab.com/gitlab-org/gitlab/-/issues/334394 for more
information.

Changelog: added