Dynamic Memory Compression: Retrofitting LLMs for Accelerated Inference