MERMaid: Universal multimodal mining of chemical reactions from PDFs using vision-language models | Publicación