Introduction of the DriveXQA multimodal dataset and MVX-LLM architecture for autonomous vehicle scene understanding in adverse conditions