What you would like to be added?

Rolling update dynamically update a PodClique's podTemplate nodeAffinity rules, without triggering pod deletions/restarts (non-disruptive update).

Why is this needed?

Training workloads may encounter machine failures. The mitigation strategy includes:

1. Automated/Manual Action: Adjust Job affinity to prevent scheduling new pods on failed nodes.
2. Pod Recovery: Recreate affected pods with strict nodeAffinity rules against faulty nodes.
3. Preservation: Unaffected pods continue running to minimize disruption.

Currently, in step1, Grove will delete all pods during rolling update, including unaffected ones.