PyramidInfer: Pyramid KV Cache Compression for High-throughput LLM Inference