SepLLM: Accelerate Large Language Models by Compressing One Segment into One Separator - View it on GitHub