Fix fractal FFT centering for odd and singleton extents - #934

Summary

Centre the fractal volume and surface filters on the actual FFT DC position for odd and single-cell extents.

The existing kernels map an index i to (i + n // 2) % n, so the DC position is n // 2. The callers currently pass a reference at n / 2. For an odd extent this introduces a half-cell offset: conjugate frequency pairs receive different multipliers. Taking the real part of the inverse FFT then changes the intended spectrum. A single-cell axis also incorrectly contributes a nonzero distance to every coefficient.

This changes the reference vectors in the volume and surface callers, including their MPI counterparts. The kernels, exponent, output shape, binning and surface-range logic are unchanged.

Regression coverage

• Odd volume extents (5, 6, 7) and (7, 7, 7), and a single-cell extent (8, 8, 1).
• Odd surface extents (5, 6) and (7, 7), and a single-cell extent (1, 8).
• The tests exercise the real caller and kernel, checking the DC reference, conjugate-pair symmetry and the single-cell-axis distance.

Verification

Using the normal source build and pytest configuration on Python 3.12:

• On the unchanged base, the seven new regression cases fail; the other 93 project cases pass.
• With the final patch, all 100 cases in test_fractal_volume.py, test_fractal_surface.py and test_fractals_generate.py pass.
• Ten additional local verification-only cases pass. Four seeded even-extent volume/surface arrays are exactly equal before and after; six odd/single-cell outputs remain finite and in range and change as expected.
• Scoped native pre-commit checks pass, including Black and isort with the repository's configured arguments.

This is not a full-repository, MPI, GPU or hardware test claim. MPI reference-vector changes were checked statically; MPI execution was not available.

Compatibility

Seeded results for odd or single-cell extents change because their spectral distances are corrected. The tested even-extent cases retain exactly the same arrays.

AI assistance

Claude Code prepared the patch and regression tests. Codex independently reproduced the defect, ran the normal-build and paired regression checks, and prepared this PR. No human maintainer approval is implied.