perf(radial): vectorize get_radial_zernikes via shared _zernike_scores (~2x)

[timtreis]

Stacked on #74 (perf/zernike-vectorize). Base will retarget to main once #74 merges. Review the radial-only diff.

What

get_radial_zernikes already built the Zernike basis on the masked foreground vectors (cheap), but reduced it with 2·K separate scipy.ndimage.sum_labels calls (one per moment × real/imag) and re-gathered pixels[ijv] inside every one. Profiling showed the reduction — not the basis — dominated.

This delegates the intensity-weighted moment sums to the shared cp_measure.utils._zernike_scores (added in #74 for get_zernike) with the pixel image as the per-pixel weight. The helper keeps the basis on the foreground vectors and segment-sums each moment by label with a single numpy.bincount, collapsing the whole reduction into one vectorised pass. The caller then normalises by each object's pixel count (radial Zernikes divide by pixel count, not the enclosing-circle area) and forms magnitude/phase.

No new dependencies. This cashes in the weight= reuse hook deliberately left on _zernike_scores.

Performance

tier	foreground	before	after	speedup
1080² medium	15.6%	315 ms	119 ms	2.6×
1080² sparse	4.1%	74 ms	42 ms	1.8×
2160² sparse	2.5%	270 ms	149 ms	1.8×
256² dense	43%	33 ms	17 ms	1.9×

Divergence vs the centrosome path: ~1e-14 (near bit-exact; the residual is summation order).

Note: this is a ~2× reduction win, not the 8× of the get_zernike PR. Unlike plain zernike, radial_zernikes never had the full-image (H,W,K) scatter — its basis was already on the masked vectors — so the win comes only from the reduction step.

Bonus: fixes a latent crash

The previous ij[label - 1] indexing assumed labels were 1..n and raised IndexError on non-contiguous label sets (e.g. {1, 3, 7}). The shared helper maps each label to its own row, so non-contiguous labels now work — covered by a regression test.