An AI operations team runs mixed training and inference workloads on a Kubernetes cluster managed with the NVIDIA GPU Operator. Inference pods frequently arrive in bursts and must start within seconds, while long-running training jobs occupy most MIG-capable A100 GPUs for days. Administrators want burst inference pods to obtain GPU capacity immediately without preempting or restarting the training jobs, and they want the cluster to reclaim those resources automatically when the burst ends. Which approach best satisfies these requirements?