[Benchmark] Update SciDocBench official release and evaluation

[Henryexcellentw]

Summary

• point SciDocBench to the official, version-pinned Hugging Face release
• materialize the portable document-image assets automatically while preserving fully offline use when they already exist
• score only the final answer after Qwen thinking and route 57 audited closed-contract tasks to deterministic Python scorers
• use GPT-5.4-mini for open-ended judging with an improved rubric and a content-addressed cache
• support both all-images-first and interleaved inputs in English and Chinese, with partition/category summaries

Dataset: https://huggingface.co/datasets/HenryExcellent/SciDocBench
Paper: https://arxiv.org/abs/2609.05141

Validation

• python -m py_compile vlmeval/dataset/scidocbench.py
• flake8 vlmeval/dataset/scidocbench.py
• 15 local scorer and portable-image regression tests passed
• verified the pinned TSV: 496 instances (124 per partition), 2,758 unique images, and no missing references

The PR intentionally changes only vlmeval/dataset/scidocbench.py; benchmark data and images remain in the official Hugging Face dataset.