Add dataset-card YAML header

Add VLMEvalKit-free runner (run_custom_model.py) + Cosmos/Qwen adapters; README

Add runnable eval code: API/local runners, GPT-judge scorer, per-task run.sh, util, README