How Attention Sinks Keep Language Models Stable