This isn't something I am planning to finish, but I wanted to see if I
would immediately run into any snags when trying to get `roachprod
start` to spin up a Docker container instead of running bare-metal.

The motivation is twofold:

- I'm generally interested in how we involve roachprod over time.
  Specifically, the question is whether in the long run systems testing
  should be done in k8s and which, if any, stepping stones there are to
  evolve what we currently do into that direction.
- The current bare-metal setup goes [unresponsive] when CRDB goes into
  overdrive, and tests fail in the most opaque ways. We then incur a
  large tax for debugging these situations since we can't access the
  cluster in that state.

Work that would need to be done to really finish this:

- maintain ubuntu images that come with docker installed (right now
  needs to be manually set up) and the `ubuntu` user set up and the images
  we need cached
- we hit the old problem of having to pass the uid/gid to the container
  to avoid creating files as root, I hacked around it by hard-coding
  them
- Setting the cluster settings, etc, is all done via the `./cockroach`
  binary but that is no longer a thing. Ideally that should use SQL
- The whole premise of uploading binaries is out of the window, we
  need to deal in CRDB containers exclusively, though we could
  conceivably use a wrapper container running the uploaded binary
  if we wanted to retain how roachprod/test work.

[unresponsive]: cockroachdb#59424 (comment)

Release note: None