[Core] Resource leakage for sky down if a multi-node cluster is partially stopped

[Michaelvll]

If a multi-node cluster is partially stopped (during autostop or manually stop the worker node), i.e. the cluster is in INIT state, our backend.teardown function will fail to terminate the stopped nodes, as we are using ray down to terminate the cluster, and ray ignores the stopped nodes.
Related code path:

skypilot/sky/backends/cloud_vm_ray_backend.py

Lines 3367 to 3369 in b962601

	elif (terminate and
	(prev_cluster_status == global_user_state.ClusterStatus.STOPPED or
	use_tpu_vm)):

To reproduce:

1. sky launch -c min --cloud gcp --num-nodes 2 --cpus 2
2. Manually stop the worker node.
3. sky down min

After that, the head node will be terminated, but the worker node is still in STOPPED state.

One possible solution is to only use our own cloud cli-based termination (should be careful: may need to kill the head node first to avoid ray autoscaler restarting some of the worker nodes, causing leakage).

[Michaelvll]

Related to this, after #2087 is merged, we should be able to move the terminate commands to the cloud class as well, making the backend.teardown function cleaner.

[HysunHe]

@Michaelvll Do we need to include the INIT status here?:

elif (terminate and
(prev_cluster_status IN ( global_user_state.ClusterStatus.STOPPED, global_user_state.ClusterStatus.INIT )