macOS CPU build + toroid phantom-collision fix (#467)

[treegardel]

Type of Change

• Bug fix (non breaking change, fixing an issue)
• New feature
• Breaking change (fix or feature that causes existing functionality to not work as expected)
• Documentation / Wiki update

Description

This PR bundles two related sets of fixes that came up while getting the macOS CPU build green.

1. macOS CPU build pipeline

• Set macOS deployment target to 13.3 and pass it consistently to CMake.
• Disable rayx-ui and CUDA only for macOS CI.
• Update CMake Apple logic so CMAKE_OSX_DEPLOYMENT_TARGET is only set as fallback (does not override workflow value).
• Add AppleClang to platform detection so RAYX_API exports symbols correctly on macOS.
• Implement macOS getExecutablePath() / getFullPath() in Locate.cpp (was static_assert(false)).
• Drop the spurious baseDir segment from the macOS resource lookup so Data/PALIK, Data/NFF, Data/CROMER, Data/MOLEC are found next to the executable (matches Linux behavior). Without this, all material tables came back empty and getRefractiveIndex returned (-1, -1).
• Fix formatAsVec redefinition on Linux (unsigned long == RandCounter).

2. Toroid phantom collisions — closes #467

The Newton solver in getToroidCollision converges to within NEW_TOLERANCE = 1e-4 mm. After a hit, the ray is nudged forward by COLLISION_EPSILON = 1e-6 mm — much less than the Newton residual. The forward check tested only the sign:

if (dot(rayToHitpoint, rayDirection) <= 0.0) return std::nullopt;

A solution within the residual on the wrong side of the new origin still passed, so each toroid bounce was counted ~4× in non-sequential tracing.

Fix: reject hits inside Newton's own uncertainty by checking parametric forward distance, local to the toroid solver:

const double t = dot(rayToHitpoint, rayDirection) / dot(rayDirection, rayDirection);
if (t <= NEW_TOLERANCE) return std::nullopt;

COLLISION_EPSILON stays at 1e-6 mm, all other surfaces keep their precision.

Tests added:

• Integration test toroid_iteration (grazing toroid vs. RAY-UI reference, regenerated at proper 1e-4 tolerance — old reference was at 1e-1 and hid the agreement).
• Unit tests on getToroidCollision: forward & backward grazing hits at z ≈ ±600 mm walking a tolerance series from 1e-2 down to 1e-6 mm (1 nm), plus a phantom-rejection regression test. Newton's quadratic convergence delivers sub-nm precision in practice.

Bonus: getCubicCollision had 29 pow(float(expr), N) calls truncating doubles to single precision on every Newton iteration — capping cubic-surface accuracy at ~1e-7 relative regardless of tolerance. Removed; all tests still pass.

This is not a breaking change. Linux/Windows workflows are not functionally changed by this PR.

Fixes #467

---

✅ Pre-Merge Checklist

• Code follows the project's coding standards
• Unit tests for new functionality are added and pass
• All existing tests pass
• Resolved TODO Comments (prefer new issues instead)
• Documentation, if applicable, including:
  • Doxygen comments for any new rayx-core API functions (N/A)
  • Helpful inline comments where needed for clarity
  • Wiki pages, e.g. updated build instructions, new Element etc. (N/A)
• Commits:
  • Use clear and readable commit messages
  • Squash and rebase onto master if individual commits don't add value
  • Ensure linear commit history (required by master)