GML-FMGroup/VisNumBench · Update dataset card for VisNumBench: Correct content, add links, sample usage, and refine tags

This Pull Request completely overhauls the dataset card to accurately describe the VisNumBench dataset.

The previous content incorrectly described a different dataset ("Deep Learning for Geometry Problem Solving"). This PR replaces it with information pertinent to VisNumBench, drawing from the associated paper, project page, and GitHub repository.

Key changes include:

Updating the primary description to reflect VisNumBench: Evaluating Number Sense of Multimodal Large Language Models.
Adding direct links to the Hugging Face paper page (VisNumBench: Evaluating Number Sense of Multimodal Large Language Models), the project homepage (https://wwwtttjjj.github.io/VisNumBench/), and the official GitHub repository (https://github.com/wwwtttjjj/VisNumBench).
Incorporating a dataset loading example for ease of use.
Refining the dataset's tags in the metadata for better discoverability, replacing the outdated geometry and mathematical-reasoning with number-sense, visual-reasoning, and benchmark.