Performance: Vectorize inner loop in _match_predictions using NumPy masking

[Vo1denz]

Problem

_match_predictions in perceptionmetrics/utils/detection_metrics.py computes
the IoU matrix in one fast NumPy call, but then reads it element by element
inside a Python inner loop over GT boxes. With N predictions and M GT boxes,
this results in N×M Python iterations per image, which becomes a significant
bottleneck at scale.

Fix

• Replaced the inner loop with NumPy boolean masking + np.argmax
• Added confidence-score-based sorting of predictions before matching
  (standard mAP practice, high-confidence predictions get priority)
• Kept used_mask as a NumPy boolean array in sync with the used set
  for fast masking

Testing

All 9 detection metric tests pass. Full suite: 44 passed, 2 failed
(pre-existing Open3D compatibility failures unrelated to this change).

I'd be happy to open a PR if this approach looks good to you.

[dpascualhe]

I've closed the PR as no actual change was provided. Feel free to open a new one. Depending on the code complexity we will decide wether we are interested in this change or not