Python 3.10–3.13 compatibility fixes + cross-version feature-equivalence test suite

[zhurong2020]

Summary

This PR makes the current master releasable on modern Python (CPython 3.10–3.13 / NumPy 2) and adds the missing scientific-trust guarantee that the modernization does not silently shift IBSI feature values.

It is a jointly-authored PR:

• @steubk's commit (fix/python312-compatibility) carries the remaining compatibility fixes on top of the already-merged scikit-build-core + NumPy 2 build migration:
  • declare ruamel.yaml (used by the CLI but previously undeclared)
  • fix deprecated scipy.ndimage.interpolation import → scipy.ndimage
  • sph_harm / sph_harm_y wrapper for SciPy ≥ 1.15
  • raise requires-python to >=3.10 (NumPy 2.1 floor)
  • add [lbp3d] optional extra (scipy / trimesh)
• My commit adds tests/version_equivalence/, a cross-version feature-equivalence regression suite.

Why

The last PyPI release (3.1.0, 2023-05) no longer installs on Python 3.12+ (see #900 / #932 / #949). master has since been modernized to scikit-build-core + numpy>=2, but a release needs confidence that the NumPy 1→2 / new-build transition reproduces established feature values — exactly what's been missing.

What the test suite checks

A deliberately broad panel to exercise the whole NumPy-2 / new-build surface:

• Image types: Original + LoG (σ=1.0) + Wavelet → C-extension filtering, SciPy LoG, PyWavelets paths.
• Feature classes: all 7 (firstorder, glcm, glrlm, glszm, gldm, ngtdm, shape) → every _cmatrices / _cshape C kernel + the pure-NumPy first-order path.
• Cases (both fully public / reproducible, no private data): a deterministic synthetic phantom (no download) and the canonical brain1 case (via getTestCase, skipped gracefully if offline).
• = 944 features × 2 cases = 1888 comparisons, each checked against a frozen reference baseline (pyradiomics 3.0.1 / NumPy 1.26.4 / CPython 3.10) at relative tolerance 1e-6.

Result

baseline	current build	comparisons	verdict
3.0.1 / NumPy 1.26.4 / py3.10	3.1.0+py312fix / NumPy 2.5.0 / py3.13	1888	✅ all within rtol 1e-6

Typical cross-stack difference is ~1e-14 (NumPy 1↔2 float reassociation on reductions such as Skewness / Kurtosis / Variance / GLCM ClusterShade) — orders of magnitude inside tolerance, the large majority bit-identical.

Notes

• "Allow edits by maintainers" is enabled.
• The build-system / compat work is @steubk's; this PR adds the verification on top of it (co-authored, not fork-and-claim).
• To regenerate the baseline under a reference env: python tests/version_equivalence/generate_baseline.py. To verify any build: pytest tests/version_equivalence -v.