Improving Visual Question Answering Models through Robustness Analysis and In-Context Learning with a Chain of Basic Questions