Extending Context Window of Large Language Models via Positional Interpolation