Fix ResourceHandle semantics

[iojw]

This fixes the following issues with the new ResourceHandle introduced in #1380.

• For clusters from the older version that are in INIT status, sky status -r will not update them to the latest version, causing the IPs for these clusters to be queried on each call which is slow. This PR updates it such that these clusters will be updated to the latest version on refresh.
• [TPU] Can't launch a TPU Pod #1480: fixes the handling of TPU pods
• The external_ips and internal_ips methods should always return a list as that is what is expected by the rest of the code

Tested:

• test_tpu_vm_pod

[Michaelvll]

Thanks for the quick fix @iojw! I am testing this with the TPU pod now.

Let's remove the following lines to make sure the TPU pod tests will be run in future smoke tests. : )

skypilot/tests/test_smoke.py

Lines 526 to 527 in 489aae1

	# Mark slow because it's expensive to run.
	@pytest.mark.slow

[Michaelvll]

Tested:

• ./tests/run_smoke_tests.sh test_tpu_vm_pod

[Michaelvll]

LGTM. Thanks for the quick fix @iojw!

[infwinston]

For smoke tests, let's try to use spot TPU pod to save some costs?

sky launch -y -c {name} examples/tpu/tpuvm_mnist.yaml --gpus tpu-v2-32 --use-spot --zone europe-west4-a

as we have some free quota in europe-west4-a.
https://github.com/skypilot-org/skypilot/pull/1481/files#diff-da9ef240243bf38ab4e33d1f8a2a64caee7677dc780e58557f3c676687a4b334R531

[iojw]

@infwinston Updated!