Efficient Streaming Language Models with Attention Sinks