Squeezed Attention: Accelerating Long Context Length LLM Inference