Fix image matching and mask helper correctness

[georgeoshardo]

What changed

• keep explicit lumMatch mean and standard deviation values independent of the number of input images
• infer masked luminance spread from the masked standard deviation rather than the masked mean
• make sfMatch finite for zero-energy radial bins and replace repeated full-image bin scans with one vectorized bin map plus bincount reductions
• preserve touching integer labels during nearest-neighbor mask resizing
• allocate floating U-Net weight maps so fractional class weights are retained
• add focused numerical regressions for each failure

Why

The helpers were losing caller-supplied luminance values, calculating one masked statistic from the wrong quantity, producing NaN/inf for constant images, merging touching instance labels, and truncating fractional weights. The Fourier implementation also repeatedly scanned every pixel once per radial bin.

Validation

• pixi run pytest -q tests/test_image_helpers.py tests/test_renderer_mask_export.py — 8 passed, 1 existing CPU-fallback warning
• pixi run pytest -q — 69 passed, 1 existing CPU-fallback warning
• old/new sfMatch comparison on 4×4, 5×7, and 32×48 inputs — maximum absolute difference at most 1.421e-13
• local full-function timing on two 256×256 images, 3 runs — median 29.7 ms before and 15.1 ms after (about 2× faster)

No benchmark test was added.

Deliberately unchanged

SyMBac/metrics.py is only imported by SyMBac/auto_optimise.py in this repository. Neither file was changed because auto_optimise.py is scheduled for removal.