DocPedia: Unleashing the Power of Large Multimodal Model in the Frequency Domain for Versatile Document Understanding