feat: add SciMIF benchmark

[shenye7436]

Summary

This PR adds support for the SciMIF benchmark in VLMEvalKit.

SciMIF contains 2,527 samples across Chemistry, Geography, Life Science, Materials Science, and Physics. It supports text-only, single-image, and multi-image samples.

Resources

• Paper: https://arxiv.org/abs/2608.25973
• Dataset: https://huggingface.co/datasets/Sheryle7436/SciMIF
• Repository: https://github.com/shenye7436/SciMIF

Changes

• Added the SciMIF dataset adapter.
• Added direct loading from Hugging Face.
• Added text-only, single-image, and multi-image prompt construction.
• Preserved edit_question and instruction_list.
• Added instruction-following evaluation.
• Added unit tests.

The current integration evaluates instruction following only. CompassVerifier-based answer-correctness evaluation is not included.

Validation

• Loaded all 2,527 samples successfully.
• Verified text-only, single-image, and multi-image inputs.
• Verified multi-image ordering.
• Verified model predictions and instruction-following evaluation.
• Tested generation model: gpt-4o.
• Tested judge model: gpt-4.1.
• Unit tests and all pre-commit checks passed.

License

SciMIF is constructed from multiple open-source scientific datasets.