[PierreRaybaut]

Closes #107

Summary

Since #104, get_radial_profile always grouped the pixels with np.unique(..., return_inverse=True) before np.bincount, so that only the bins holding pixels were allocated. The sort it involves made radial profiles of large images 2 to 3 times slower than in Sigima 1.3 in the usual case (uniform coordinates, normal step), where the bins are far fewer than the pixels.

• get_radial_profile now chooses the grouping strategy from the number of bins, with no tuning:
  • dense bins (largest bin index lower than the number of pixels, i.e. the usual case): direct np.bincount on the bin index, empty bins left out as before — no sort, no more entries than pixels. The bin index is converted to integers only after this test, since it is unbounded for tiny steps;
  • sparse bins (tiny steps): unchanged np.unique + np.bincount path.
• The memory used still never exceeds the size of the image, whatever the step, and results are bit-identical in both paths (same bins, same order of accumulation).
• The contract of radial_profile is unchanged (version 2). No release note: the slowdown was never released (Define image profiles in physical coordinates #98 and Allocate only the occupied bins of radial profiles #104 are later than v1.3.0).

Performance

Whole get_radial_profile, random image, off-pixel center, step 1, NumPy 2.4.4 (best of 7):

Image	v1.3.0 (direct bincount)	develop (np.unique)	This PR
1000 × 1000	0.023 s	0.050 s	0.024 s
2000 × 2000	0.095 s	0.215 s	0.100 s
4000 × 4000	0.380 s	0.974 s	0.395 s

The small remaining difference with v1.3.0 comes from the input checks added by #104.

Validation

• New tests (profile_physical_unit_test.py):
  • bit-identity with the former implementation now also covers sparse bins (tiny gap between two pixel centers, about 10⁵ bins for 300 pixels), the existing cases covering dense bins (float, integer and masked data);
  • guard against the slowdown: np.unique is not called for dense bins, and is called for sparse bins (this test fails on develop).
• Full suite: 2366 passed, 63 skipped; --qualification-report: no problem (radial_profile v2 qualified, validation and replay exact).
• DataLab develop against this branch: profile_app_test.py, provenance_processing_unit_test.py, interactive_processing_test.py pass (52).