Add BVH acceleration for mesh intersection in CUDA path tracer

[jichumao]

Motivation

• Reduce expensive per-mesh brute-force triangle traversal during ray intersections by adding a CPU-built BVH per mesh and using it in device intersection code.
• Provide data structures and device buffers to upload a flattened BVH to CUDA so the path tracer can traverse the hierarchy on the GPU.

Description

• Introduce BVH data structures and per-geom root index by adding bvhRootNodeIdx to Geom and a compact BVHNode layout suitable for device use (src/sceneStructs.h).
• Add scene-level storage for flattened BVH arrays (std::vector<BVHNode> bvhNodes and std::vector<int> bvhTriIndices) and build a per-mesh BVH during glTF loading (src/scene.h, src/scene.cpp).
• Replace linear triangle scanning with BVH stack traversal in meshIntersectionTest, and replace the previous AABB test with a robust slab-based implementation; update the function signature to accept BVH buffers (src/intersections.h, src/intersections.cu).
• Upload BVH buffers to device memory and wire them into the intersection kernel and computeIntersections call, and free them on shutdown; enable BVH_ENABLED by default (src/pathtrace.cu, src/utilities.h).

Testing

• Attempted a full build with cmake -S . -B build && cmake --build build -j4, which failed due to the environment lacking the CUDA toolkit / nvcc (so compilation could not be validated here).
• No other automated tests were present or executed in this environment.

---

Codex Task