Optimize Metropolis2D IoU evaluation

[raus7n]

Summary

This PR optimizes the Metropolis2D 2D detection evaluation by precomputing the prediction-to-ground-truth IoU matrix once per image and reusing it across all IoU thresholds.

Motivation

The evaluator uses 10 IoU thresholds from 0.50 to 0.95. Previously, the same prediction/ground-truth IoU pairs were recomputed independently for each threshold.

For N predictions, M ground-truth boxes, and T IoU thresholds, this results in approximately:

O(T × N × M)

IoU calculations.

Changes

• Precompute the N × M IoU matrix once for each image.
• Reuse the precomputed IoUs during matching across all IoU thresholds.
• Preserve the existing confidence ordering and greedy matching logic.
• Preserve the existing per-threshold ground-truth matching behavior.
• Leave the AP/mAP calculation unchanged.

This reduces the number of IoU calculations from approximately:

O(T × N × M)

to:

O(N × M)

where T = 10.

Validation

• python -m py_compile evaluation/cosmos3/reasoner/vlmevalkit/vlmeval/dataset/metropolis2d/detection_2d_dataset.py
• git diff --check

Testing

The existing evaluation logic and matching semantics are unchanged. The optimized implementation was syntax-checked successfully.