Can LLMs Maintain Fundamental Abilities under KV Cache Compression?