Identify Critical KV Cache in LLM Inference from an Output Perturbation Perspective