Support full Euler angles (alpha, theta, psi) in parallel Siddon projector

[hashimoto-koh]

Description

For geo.mode == "parallel" with projection_type == "Siddon", the projector previously only applied 2D rotation around the Z-axis using geo.alpha, completely ignoring geo.theta and geo.psi.

This prevented parallel-beam laminography and tilted-axis geometries from being simulated with the Siddon projector.

This PR introduces the 3D ZYZ Euler rotation (eulerZYZ) into Common/CUDA/Siddon_projection_parallel.cu, ensuring that all three view-dependent angles (alpha, theta, psi) are correctly applied to the parallel rays and detector plane according to TIGRE's geometry convention.

In addition, it completes the 3D Siddon kernel by:

1. Restoring Z-axis plane stepping and ray bounds (azm, azM, kmin, kmax, az, azu).
2. Ensuring the plane traversal invariant ($a > a_m$) and strictly determining initial voxel indices from ray direction and boundary planes, preventing voxel misidentification and segment dropouts caused by single-precision rounding on grazing/cardinal rays.
3. Preserving bitwise consistency with upstream baseline for untilted parallel projections (theta=0, psi=0) by retaining original division operators.
4. Returning early for rays that do not intersect the volume AABB (am >= aM), which is required for correct ray tracking in tilted geometries.

Verification

• Sensitivity: Verified that varying theta and psi independently alters projections of an asymmetric 3D volume as expected.
• Untilted Baseline Consistency: Verified across 11 arbitrary alpha views (11,264 detector pixels) on a random 32^3 volume that theta=0, psi=0 matches upstream master baseline with zero error (max absolute error = 0.0).
• Independent Analytical Ground Truth: Verified ray path lengths against an independent physical-coordinate analytical Ray-Box intersection calculation across 13 views with arbitrary 3D Euler angles, anisotropic voxels, and offsets (max error < 0.02, correlation > 0.99999).
• Grazing and Boundary Rays: Verified on grazing rays (theta=pi/2) and cardinal entry views (72 views probe) that marcher integration matches actual GPU endpoint Ray-AABB intersection lengths within ~5.4e-5. (Note: as with existing float32 projectors, grazing rays extremely close to voxel boundaries are subject to single-precision coordinate construction rounding).
• Regression Tests: Added Python/tests/test_parallel_siddon_euler.py containing 7 comprehensive regression test cases (all passing).