This PR improves the dataset card by:
- Linking the dataset to the original research paper: [MM-SpuBench: Towards Better Understanding of Spurious Biases in Multimodal LLMs](https://huggingface.co/papers/2406.17126).
- Adding the `arxiv` tag to the YAML metadata.
- Updating the `task_categories` to include `image-text-to-text` to better reflect the multimodal nature of the benchmark.
