Layer-Condensed KV Cache for Efficient Inference of Large Language Models