26× Faster Inference with Layer-Condensed KV Cache for Large Language Models