refactor: GPU model interface API

[tobre1]

Title: Refactor GPU model interface API and add tracing benchmark

Breaking change: This changes the GPU model interface API. Custom GPU models and pipelines must update collision and reflection callable signatures to accept an explicit primitive ID, adopt the new shader names, and adjust access to per-ray data.

• Move disk neighbor intersections into a separate payload to reduce shared per-ray state.
• Read traversal data directly from OptiX and advance ray positions in closest-hit shaders.
• Add a GPU benchmark for disk and triangle geometries with accompanying datasets.