Fix AP/mAP degradation due to unsorted prediction confidence scores & Add IoU Zero-Division Safeguard

[shardulmahamulkar]

This PR addresses two critical bugs in the perceptionmetrics detection evaluation pipeline that result in artificially degraded AP (Average Precision) scores and potential crashes:

Prediction Confidence Sorting (#Bug 3): Standard evaluation mapping protocols dictate that multiple boxes targeting a single ground truth must be resolved sequentially in descending order of their confidence scores. The _match_predictions() method was evaluating mapping criteria based on arbitrary array index order. This allowed low-scoring detections to lock in early as the True Positive target, delegating higher confidence detections as False Positives, unjustly penalizing high-threshold precision curves.

Degenerate Area Bug / IoU Zero-Division Safeguard (#Bug 1): compute_iou() did not protect against zero-area unions (i.e., evaluating degenerate predicted or ground-truth boxes), causing ZeroDivisionError crashes mid-evaluation. A safeguard was implemented to return 0.0 seamlessly in these cases without halting execution.

Changes Made

• Added np.argsort() within perceptionmetrics/utils/detection_metrics.py to correctly map predictions traversing from highest confidence to lowest.
• Protected Python float division in compute_iou() against division-by-zero crashes, utilizing if unionArea > 0 else 0.0.

How to Verify

• Run detection evaluation against overlapping low-confidence/high-confidence objects—AP/mAP calculations will accurately penalize the duplicates and effectively reward the peak detection thresholds.
• Execute compute_iou() on overlapping degenerate/empty boxes—the application correctly resolves an IoU of 0.0 instead of crashing.

Related Issues

Closes #544