Track DRA/NRI allocation lifetime assumptions before future sharing or preemption support

[AutuSnow]

StopContainer currently releases claim CPU allocations before UnprepareResourceClaims runs.

The code comments already document this as a workaround and note that it assumes ResourceClaims are not shared across pods/containers. That is fine for current behavior because claim sharing is explicitly unsupported.

This issue is to track the assumption before future work changes the lifecycle model.

Related:

• KEP-5690: Add proposal (DRA preemption) kubernetes/enhancements#6113
• KEP-5981: DRA Sharing Affinity for Consumable Capacity kubernetes/enhancements#5987
• Support mixed shared and isolated CPU allocation with CPU DRA #103
• [Do Not Merge] POC for Mixed CPU Allocation #149
• Support NodeAllocatableResourceMappings for DRA CPU scheduler accounting #179

Suggested:

• document the current DRA/CDI/NRI lifetime invariant
• add tests around StopContainer, UnprepareResourceClaims, and restart Synchronize
• revisit this before enabling claim sharing, preemption, or mixed allocation behavior