[Benchmark] Support ScienceOlympiad Galaxy10DECaLS VRSBench

[zhouyujin]

Add three datasets: ScienceOlympiad, Galaxy10DECaLS, VRSBench

1. ScienceOlympiad
   TSV link: https://huggingface.co/datasets/YuJJJJin/ScienceOlympiad.tsv
   ScienceOlympiad focuses on competitive‑level physics and chemistry problems with multimodal content. It evaluates models on scientific reasoning and visual comprehension.
2. Galaxy10DECaLS
   TSV link: https://huggingface.co/datasets/YuJJJJin/Galaxy10DECaLS.tsv
   Galaxy10DECaLS is a curated image classification dataset with 1,774 galaxy images across 10 classes. It evaluates models’ ability to classify astronomical objects based on visual features.
3. VRSBench
   TSV link: https://huggingface.co/datasets/YuJJJJin/VRSBench.tsv
   VRSBench is derived from the VQA test set of the VRSBench benchmark and evaluates multimodal understanding of remote‑sensing imagery.
   Two variants are provided:
   • VRSBench.tsv: Full evaluation set with 37,409 VQA samples.
   • VRSBench_MINI.tsv: Compact evaluation set with 3,735 samples (10% stratified sampling from the full set, seed=42).
   Both datasets cover 12 question categories and assess a model’s ability to answer remote‑sensing questions through visual analysis and reasoning.