Align imaging orientation to Allen RAS+ + transform helpers + viz contrast fix

[FIrgolitsch]

Summary

Extends linumpy.imaging to align 3D volumes to the Allen RAS+ template convention, and fixes contrast scaling in the orthogonal-view renderers. Changes the orientation target from (S, R, A) to (S, A, R) so a reoriented volume feeds directly into numpy_to_sitk_image as a SITK image ordered (X=R, Y=A, Z=S) — matching the Allen template and avoiding a spurious 90° rotation that the rigid registration was failing to recover.

Land after #140.

What changed

• orientation.py — convention change. parse_orientation_code now targets (S, A, R) instead of (S, R, A). The letter_map and all docstring/examples are updated. This changes the permutation/flip output for every input code, so any caller caching or asserting on a specific permutation needs updating.
• orientation.py — new helpers for downstream RAS alignment:
  • sitk_transform_to_affine_matrix — Euler3D / AffineTransform → 4×4 matrix in (Z, Y, X) OME-NGFF ordering.
  • store_transform_in_metadata — writes that affine as an OME-Zarr coordinateTransformations block.
  • compute_centered_reference_and_transform — builds a zero-origin reference grid and the composite transform for resampling a moving image into it.
• visualization.py — contrast fix. Orthogonal and annotated views now compute a per-panel vmax from positive, finite voxels only (percentile) and clamp vmin = 0, instead of a single global min/percentile over all panels. Fixes washed-out previews when one panel is dominated by zero/negative background.
• Tests. test_imaging_orientation.py re-targeted to the (S, A, R) convention (identity is now SAR, PIR/RAS/LPS expectations updated). New test_imaging_orientation_ras.py covers the three new helpers (identity/translation permutation, brain-volume preservation under rotation, affine-block round-trip into .zattrs).

Reviewer focus

• The convention change is load-bearing and subtle. It exists so the output matches the Allen template handed to register_3d_rigid_to_allen (next PR). The rationale is in the module docstring; the updated PIR/RAS/LPS test expectations are the best place to confirm the new mapping is correct.
• compute_centered_reference_and_transform derives the output grid size from the transformed corner points — verify the rounding (np.ceil) matches how downstream resamplers consume it.
• store_transform_in_metadata prepends an affine to each dataset's existing coordinateTransformations; check the ordering is right for your reader.

How to verify

uv run ruff check linumpy/imaging/
uv run ty check linumpy/imaging/
uv run pytest linumpy/tests/test_imaging_orientation.py linumpy/tests/test_imaging_orientation_ras.py -x -v