Run on existing tpu vms (on demand - not queued)

[kathir-ks]

The existing logic is to setup a queued resource and wait for it to be resourced by google cloud to start training.
But what if the user already has the tpu vm ready with him?
May need to add few changes in the launch.py and tpu.py.

[dlwh]

it's a good point! queued-resources is necessary for multislice which is why we prefer that. You can use infra/spin-up-vm.sh (which will skip spinning up and instead initialize a venv if it exists)