Support kaniko - incluster builds

[jlewi]

Now that we have code for building the context launching as a Job running kaniko in K8s should be relatively straightforward.

Possible advantages of using K8s

• Pools of available resources
• More resources (in particular RAM) than what's available in GCB

This PR has some YAML for launching kaniko on k8s
jlewi/notes#12

[jlewi]

Not sure how useful this would be.
It looks like kaniko has a dependency on ephmeral-storage GoogleContainerTools/kaniko#2219. I think on GKE clusters this maxes out at 10Gi on standard clusters and 1Gi on Autopilot clusters. This turns out to be too low for really big images like CUDA.