A benchmark multimodal oro-dental dataset for large vision-language models