Minimize base cost of our clusters

[yuvipanda]

I want our clusters to be as cheap as possible when nobody is using them. This means:

1. Just one core node kept up at all times, and this should be as small as possible.
2. Autoscale up only when users start it - we could either have one user node running, or 0 as we see fit. We can actually have enough room in the core node for just a placeholder pod, and then have a new node come up when the first user shows up.
3. Figure out what to do with the NFS server - probably Run NFS servers in-cluster #50
4. Run zonal clusters, not regional ones.

[yuvipanda]

The smallest cluster you can run needs at least 4 cores, which is atrociously high. Almost all this is kube-system pods, and should really be counted as the GKE management fee. https://cloud.google.com/kubernetes-engine/docs/how-to/small-cluster-tuning has some suggestions, but isn't good enough. https://www.eficode.com/blog/custom-dns-on-gke also has useful information.

n1-custom-4-3840 is the smallest core node we can get. n1 is preferable over e2 since these nodes will be on all the time. It'll be 99% idle most of the time, which is infuriating, but hey that's the tradeoff for having a managed service!

I also performed some 'bring down' actions:

• Make NFS VM smallest possible
• Move the home directories away from their own external disk to the NFS VM's base disk.

NFS should definitely go in-cluster - see #50

[yuvipanda]

#2199 is a better version of this.