Why are these changes needed?

Without this PR, RayJob has a complex behavior to pass changes from RayJob's spec to the associated RayCluster.

• If Ray Autoscaling is enabled, do not pass any changes to the associated RayCluster.
• If the change concerns only replicas, update the RayCluster.
• Otherwise, do not update the RayCluster.

This PR decides to remove the behavior because:

• For batch jobs, the required computation resources should be deterministic. Hence, no need to pass changes to RayCluster.
• If users still wish to update replicas at runtime, they should use the Ray Autoscaler instead.
• If users still want to update replicas manually, they still can update RayCluster directly.
• Allowing users to update replicas at runtime will make gang scheduling much more complex.
• The best practice for batch jobs should include the ability to tolerate downtime and implement application-level fault tolerance within the jobs themselves, such as writing and reading checkpoints from S3.

Hence, I remove this behavior in this PR to avoid unnecessary complexity of the RayJob codebase.

Related issue number

Checks

• I've made sure the tests are passing.
• Testing Strategy
  • Unit tests
  • Manual tests
  • This PR is not tested :(