Extending Context Window of Large Language Models from a Distributional Perspective