[Dataset] Add BenchCAD (CAD code-understanding VQA)

[HaozheZhang6]

What

Adds BenchCAD, a benchmark for evaluating MLLMs on mechanical CAD parts, with two configs:

Dataset	Input	Output
BenchCAD_VQA	rendered views of an industry-standard part	a single number
BenchCAD_CodeQA	rendered views + CadQuery source	a single number

2,400 questions over 200 parts / 106 part families / 49 engineering standards (dimensions, counts, ratios).

Scoring

Deterministic, no judge model required:

• exact match for counts / integers / yes-no
• symmetric ratio accuracy min(pred, gt) / max(pred, gt) for dimensions and ratios

evaluate() reports an overall score plus a breakdown by question type. Mirrors the official BenchCAD scoring (https://github.com/BenchCAD/BenchCAD-main).

Data

Single TSV hosted on HuggingFace: BenchCAD/BenchCAD → vlmevalkit/BenchCAD_QA.tsv (md5 e19747288563fe807d0ceb8049da1989), images base64-encoded inline.

Testing

• flake8 / isort / yapf (column_limit=120) all clean on the new file and the registry edit.
• Unit-tested the scoring (_to_number, _score_one) and evaluate() aggregation against hand-computed values (overall/integer/dim/ratio), and build_prompt for both configs (VQA omits code, CodeQA includes the CadQuery block).

Notes

BenchCAD is described in arXiv:2605.10865 and is featured as an external benchmark in Anthropic's Claude Opus 4.8 system card (§8.16.4). Follows the ChartQAPro dataset pattern; one new file vlmeval/dataset/benchcad.py plus registry import/list entry in vlmeval/dataset/__init__.py.