Fix Siddon projection kernel missing return on missed rays (cone and parallel)

[hashimoto-koh]

Description

In Siddon_projection.cu and Siddon_projection_parallel.cu, when a ray does not intersect the bounding box of the voxel volume (am >= aM), detector[idx] is set to 0, but the kernel forgets to return.

Consequently, GPU threads for missed rays continue executing subsequent index computations, texture fetches, and loop iterations with invalid index ranges (such as imin > imax), causing unnecessary GPU instruction overhead and potential undefined behaviors.

This PR adds an early return; immediately after zeroing the detector pixel on missed rays in both cone-beam and parallel-beam Siddon projection kernels.

Changes

• Common/CUDA/Siddon_projection.cu: early return when am >= aM
• Common/CUDA/Siddon_projection_parallel.cu: early return when am >= aM
• Python/tests/test_siddon_missed_rays.py: regression test for missed rays in both cone and parallel modes

[AnderBiguri]

Ca you clarify how this PR is different to this one: #779?

[hashimoto-koh]

Hi @AnderBiguri,

Apologies for the oversight — I hadn't noticed #779 when opening this PR.

You are completely right: the core fix (adding the missing return; when am >= aM in both Siddon_projection.cu and Siddon_projection_parallel.cu to prevent infinite looping / GPU hang) is essentially identical to #779.

The main differences are:

1. No binary files: This PR does not include the PNG image asset or figure generators present in Siddon: a missed ray must RETURN, or the projector never comes back #779.
2. Minimal & fast regression test: This PR includes a small, lightweight unit test (Python/tests/test_siddon_missed_rays.py) that runs in under a second without multiprocessing/timeouts, and passes all CI checks.

Next steps (up to your preference):

• If you prefer to wait for the author of Siddon: a missed ray must RETURN, or the projector never comes back #779 to address your comment (removing the image and tests) and merge their PR to credit them, please feel free to close this PR.
• Alternatively, if you would prefer to move forward here (or if you would like me to remove the test file as well to keep only the CUDA changes, as you suggested in Siddon: a missed ray must RETURN, or the projector never comes back #779), I am happy to update this PR accordingly.