fix(scheduler): namespace PodGroup identities - #2095

Description

PodGroup is a namespaced Kubernetes resource, but the scheduler cache used only PodGroup.Name for three related identities:

• the pod informer index,
• the snapshot's PodGroupInfos map, and
• the job ID assigned to pods in the snapshot.

As a result, same-named PodGroups in different namespaces collided. One snapshot entry overwrote the other and pods from both namespaces could be attached to one group. This could make topology allocations owned by one workload appear to belong to another workload and incorrectly remove otherwise available nodes from consideration.

This change introduces a canonical namespace/name PodGroup ID and uses it consistently at the cache boundary. Pods without a PodGroup annotation retain the existing empty index key.

A regression test creates live/shared-name and shadow/shared-name with one pod each. Before the fix, the test observes one merged group; after the fix, it observes two isolated groups whose pods carry the correct namespaced job ID.

Related Issues

N/A

Checklist

• Self-reviewed
• Added/updated tests
• Documentation not required (internal cache identity fix)
• Added a changelog fragment via make changelog

Breaking Changes

None. The change affects internal scheduler cache keys only.

Additional Notes

Validation on an x86_64 Linux workstation:

• go test ./pkg/scheduler/... -count=1
• go vet ./pkg/scheduler/...
• golangci-lint run -v -c .golangci.yaml ./pkg/scheduler/...
• go build -buildvcs=false ./cmd/scheduler

The regression is deterministic cache bookkeeping and does not require a multi-node GPU cluster.