MiniCache: KV Cache Compression in Depth Dimension for Large Language Models