[serve] Make serve.shutdown() shut down remote Serve applications

[shrekris-anyscale]

Why are these changes needed?

serve.shutdown() does not do anything if there is no cached handle to a Serve controller. If a user connects to a remote Ray cluster without running serve.start(), they cannot use serve.shutdown() to shut down the remote Serve application.

E.g. this does nothing, even if a Serve controller is running at "cluster_address" in namespace "x":

import ray
from ray import serve

ray.init(address="cluster_address", namespace="x")
serve.shutdown()

This change makes serve.shutdown() first try to connect to a Serve controller before attempting to shut down. This allows a user to directly shut down a running Serve instance without potentially having to start a Serve instance with serve.start() first.

Related issue number

Closes #22254.

Checks

• I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://docs.ray.io/en/master/.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
• Testing Strategy
  • Unit tests
    • New unit tests are added to test_standalone2.py.

[edoakes]

@shrekris-anyscale there was a windows test failure in test_standalone. Retried, but can you verify if it's a real failure?

[shrekris-anyscale]

@edoakes It is a real failure– Windows has some issues when trying to run temp files. I can either skip the test on Windows or refactor it to pass. I'll look into this.

[shrekris-anyscale]

@edoakes I removed the context manager from the test to handle the Windows issues (see https://stackoverflow.com/a/15590253 for context), and it looks like test_standalone2.py is passing in Windows and Linux now.

[edoakes]

@shrekris-anyscale can we update the DeploymentNode docstring to say "deployment graph" instead of "multi-deployment applicaton" ?

[edoakes]

Sigh... wrong PR, ignore above. Clearly reviewing while distracted today...