Integrate face anonymization for photogrammetry scans

[harmening]

Integrate face anonymization for photogrammetry scans (builds on #166)

Builds on @Mah-diaa's #166, which contributed the whole face-anonymization algorithm and its test suite. Thanks a lot for that work!
This PR mostly wires it into cedalion's conventions and tightens a few edge cases; the method is unchanged.

What changed vs #166

• Module layout. pipeline.py is now a thin anonymize_scan orchestrator, shared helpers are consolidated in _utils.py, all mask logic (including the landmark carve-outs) lives in mask.py, and preprocessing owns the framing steps.
• CRS/units integration. The pipeline stays in xarray + pint and reuses the existing points.apply_transform and affine_transform_from_numpy. Transforms carry their CRS: revert_to_einstar_frame takes an AffineTransform and derives the target frame from it instead of a hardcoded string. anonymize_scan validates that surface and landmarks share the same CRS.
• API consistency. Renamed normalize_axes -> orient_y_anterior and align_axes_from_landmarks -> align_to_ctf, 4×4 transforms throughout, and cap parameters grouped into a frozen CapDetectionParams.
• Broader inputs + guards. Works also on textureless meshes now; fixes cap detection when the profile peaks at the first bin; head isolation warns instead of silently keeping the body; ValueError on degenerate landmarks (LPA==RPA, Nz on the ear axis); RuntimeError if masking removes most of the mesh.
• Tests + docs. Expanded the tests (including the CRS-mismatch and textureless paths) and added a Photogrammetry docs page.

Original implementation: @Mah-diaa (#166).