Generalizing Visual Question Answering from Synthetic to Human-Written Questions via a Chain of QA with a Large Language Model