GitLab

Instead of giving up when a single datastore fails, operate on all datastores
first. That way:

* If one datastore doesn't support disk usage, we can at least return our best
  estimate for the ones that do.
* If one datastore fails to sync, we won't fail to sync the others.
* If one datastore refuses to close, we'll close the others.

Etc.

Also:

* Drop xerrors. We already need go 1.13 elsewhere.
* Use uber's multierr. It's much simpler than hashicorps (and is already a
  dependency through zap, fx, etc.).