GaussianProjectionForward: fix camera data loading that exceeds blockDim

[swahtz]

In loadCamerasIntoSharedMemory, each thread loaded an element of camera data into shared memory. However, if the amount of camera data exceeds the number of threads, all of the camera data would fail to be loaded. This happens around camera 13 (256 / 21 elements-per-camera). Changed this loading into a strided loop so all camera data will be loaded.

Also I changed the calculation of shared memory from a manual calculation to using sizeof() the appropriate nanovdb::math structs for safety's sake.

[Copilot]

Pull Request Overview

This PR fixes a critical bug in the Gaussian projection forward operation where camera data loading would fail when the number of cameras exceeded approximately 13 (256 threads / 21 elements per camera). The fix implements a strided loop pattern to ensure all camera data is loaded into shared memory regardless of the number of cameras, and improves the safety of shared memory calculations by using sizeof() operators.

• Replaced single-pass conditional loading with strided loop to handle camera counts exceeding block dimensions
• Changed shared memory type from T[] to char[] and updated pointer arithmetic to use sizeof() for type-safe offset calculations
• Made shared memory size calculations more explicit and type-safe using sizeof(nanovdb::math::Mat3<scalar_t>) and sizeof(nanovdb::math::Vec3<scalar_t>)

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[harrism]

Nice fix! Just think maybe the shared declaration should be aligned as in your other recent PR.