Non-Maximum Suppression

A detector doesn't find one box per object — it finds a dozen nearly identical ones. NMS is the greedy cleanup that keeps only the best of each pile:

1. Take the highest-scoring box that's still in play and keep it.
2. Discard every remaining box whose IoU with it exceeds iou_threshold — those are duplicates of the same object.
3. Repeat with what's left.

Task: write nms(boxes, scores, iou_threshold) where boxes is a list of [x1, y1, x2, y2] and scores is the matching confidence for each. Return the list of kept indices, in the order they were kept (so highest score first).

• Boxes are removed only when IoU is strictly greater than the threshold — exactly equal survives.
• If two boxes tie on score, the lower index goes first.
• Empty input returns an empty list.

Note what the threshold does. Set it high and you keep near-duplicates; set it low and you start deleting genuinely separate objects that happen to stand close together.