Fix evaluate_from_hf.py: un-pivot columnar annotations, load parquet directly

Sync evaluate_from_hf.py: filter redundant/decomposable boxes and stable-sort ground truth to match loading.load_annotation_boxes

Add requirements.txt, fix split default

Add standalone evaluation kit