[metrics] Raise a clear error for self-intersecting ground truth boxes - #2219

Request changes

Thanks for tracking this down. Crashing deep inside GEOS with a TopologyException is a bad experience, and the fix itself works. But I don't think we should quietly repair boxes inside polygon_iou.

A ground-truth box whose points cross each other is a broken annotation. If we reorder its points without telling anyone, the metric reports a score against a box that isn't in the dataset, and the annotation error is hidden. Rather than accept wrong ground truth, we should reject it with a clear error so the dataset gets fixed.

Requested changes:

1. Remove the automatic reordering (_polygons / order_points) from polygon_iou.
2. Raise a clear ValueError instead of the GEOS exception. It should name the offending boxes, e.g. "Ground truth boxes at indices [3, 17] are self-intersecting (points not in TL, TR, BR, BL order); please fix the annotations." Since polygon_iou can't tell ground truth from predictions, the cleanest place for this is a small helper called on gts in the update() methods that use use_polygons (LocalizationConfusion, OCRMetric, DetectionMetric, ObjectDetectionMetric, TableCellMetric). If you think predictions should be validated the same way, I'm open to that too.
3. Don't reject degenerate boxes. Zero-area boxes (collinear points, repeated points) also fail is_valid, and collinear ones even report Self-intersection. Main returns IoU 0 for these today and that should stay the same. Only flag boxes that are invalid and have a non-zero convex-hull area, i.e. a real box whose points are out of order.
4. Tests: replace test_polygon_iou_misordered_points with tests that check:
   • a crossed ground-truth box raises ValueError with its index in the message;
   • a batch mixing good and bad rows reports only the bad indices;
   • zero-area / collinear boxes still give IoU 0 with no error.

On performance: I benchmarked the is_valid check against main, and it adds about 2% end-to-end on LocalizationConfusion with polygons, so a vectorized validity check is fine to keep.

Thanks 👍