Parameter-Free KV Cache Compression for Memory-Efficient Long-Context LLMs