Speed up FROC curve computation #9231

compute_froc_curve_data compares every detection against every distinct score, so it's O(N²) when scores are mostly unique. Sorting once and counting with searchsorted gives the same arrays and is much faster (17× at 11k detections, 55× at 55k, single-thread CPU).

i check and outputs matched exactly in 328 cases on NumPy 1.26 and 2.4. When all scores are equal it's a hair slower (0.03 ms). Happy to send a PR if this is welcome.

also happy to send over the test sscripts in that PR.

Disclosure: I used Pi with Claude Opus and GPT-6 for profiling and benchmarks.