Use preemptible node pool for runners

When installing GitLab under Kubernetes using our bundled Helm chart, we give no guidance to the type of cluster they should use. Google's GKE supports preemptible VMs at much cheaper rates. We should explore and come up with a recommendation for using them. Some best practices:

• We should have a fixed, regular node pool, with affinity for GitLab Omnibus, etc.
• We should have an autoscaling, preemptive node pool with affinity for Runners

Note GKE's beta support for autoscaling doesn't take cost into account, so to control costs, only one pool should be autoscaling.

Having some Runner capacity in the regular node pool can help keep things moving in the event of preemptive VM contention.

Links

• https://cloud.google.com/container-engine/docs/preemptible-vm
• https://cloud.google.com/container-engine/docs/cluster-autoscaler