Streamlining Redundant Layers to Compress Large Language Models