Why are these changes needed?

When ray.get(master_actor.create_backend.remote(backend_tag, backend_config, replica_config)) is called by serve.create_backend, this raises an RayActorError exception because master_actor just literally killed. At the release test, we don't catch this exception, so it crashes the driver (because there's uncaught exception). This kills Raylet & GCS server, which introduces a weird errors in there.

I don't know what's the right behavior in this case, but it seems like wrapping serve.create_backend, serve.create_endpoint, serve.delete_endpoint, serve.delete_endpoint by

try: 
except ray.exceptions.RayActorError:

resolves the issue.

Conclusion: the problem is that we don't catch RayActorError anymore, and that crashes the driver. In the past, we manually caught RayActorError to retry actor task, but not anymore after we rely on Ray's actor retries mechanism. related PR: 4155d58)

##NOTE
This is a hotfix, and I am not sure if it is the right behavior you want. Feel free to reject the PR if we need other solutions.

Related issue number

Closes #8915

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/latest/.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failure rates at https://ray-travis-tracker.herokuapp.com/.
• Testing Strategy
  • Unit tests
  • Release tests
  • This PR is not tested (please justify below)