Having an invalid KUBECONFIG without a token in it is not helpful. This
only became possible recently now that we are creating a separate
namespace and service account (and hence token) to send to the runners.
This led to somewhat surprising results when troubleshooting
https://gitlab.com/gitlab-org/gitlab-ce/issues/53879 as I found that the
KUBECONFIG was still being passed but KUBE_TOKEN was not. These things
really should have been linked.

Furthermore now that we are also using the [presence of KUBECONFIG to
decide whether or not to run build steps in Auto
DevOps](https://gitlab.com/gitlab-org/gitlab-ce/blob/294d15be3e9497e7b67e1f9131ce9d5c0d68406c/lib/gitlab/ci/templates/Auto-DevOps.gitlab-ci.yml#L164)
I think it makes even more sense to ensure that KUBECONFIG is a complete
config if passed to a job.