Store or Recompute? Characterizing the Carbon Tradeoff of KV Cache Retention in LLM Inference