Exploiting Local KV Cache Asymmetry for Long-Context LLMs