In CI we don't have any replicas set up. In this case retrying any
connection errors (e.g. when a job runs without a DB set up) can
increase the job timings.

To fix this, we simply don't retry connection errors in case no replicas
are used. This also ensures the retry behaviour is the same as before
the introduction of always enabling the load balancing code.

Finally, we no longer treat NoDatabaseError errors as connection errors.
There's no point in retrying these errors, even if replicas are present,
as retrying isn't magically going to make your database exist.

See https://gitlab.com/gitlab-org/gitlab/-/issues/341757 for more
information.

Changelog: fixed