Extract Information from Hybrid Long Documents Leveraging LLMs: A Framework and Dataset