[NV] Add LiDAR sensor support for 3D Gaussian Splatting rasterization

[rodlima78]

---

Summary

• Lidar camera model: Introduces a spherical-coordinate lidar sensor with configurable elevation/azimuth beams, supporting both uniform and non-uniform beam distributions, plus rolling-shutter pose refinement.
• Lidar tiling and intersection: Adds a dedicated tile-based intersection pipeline (IntersectTileLidar.cu) that partitions the spherical field of view into azimuth/elevation tiles and assigns Gaussians to
  tiles based on their projected 2D covariance ellipses.
• Lidar rasterization: Extends the rendering pipeline (rendering.py, _wrapper.py) to support lidar sensors end-to-end — projection, tiling, alpha-compositing — reusing the existing 3DGS/3DGUT rasterization
  kernels with distance-based output.
• Python reference implementation: Adds pure-PyTorch lidar equivalents (_torch_impl_lidar.py, _torch_lidars.py) for CPU testing and cross-validation against the CUDA kernels.
• Bug fixes and robustness: Fixes singular matrix inversion in SH evaluation, corrects hit-distance clipping when global_z_order==false, adds a torch.tensor_hash fallback for PyTorch < 2.9, and removes
  premature culling of small Gaussians.
• Comprehensive test coverage: Adds extensive tests for lidar projection, tiling, intersection, and rasterization, as well as a new 3DGUT distance-output rasterization test.

Key details

• Lidar beam layout is defined by per-beam elevation/azimuth angles plus a beam-to-pixel mapping, allowing irregular scan patterns (e.g., Velodyne, Ouster, Hesai).
• Tiling operates in spherical coordinates — azimuth bins form columns, elevation bins form rows — with special handling for azimuth wraparound across the 0°/360° boundary.
• Rolling-shutter support iteratively refines Gaussian positions using per-column ego-motion transforms, controlled by rs_num_iters.
• global_z_order=false switches rasterization from depth-based to hit-distance-based ordering, which is the natural sorting for lidar returns.
• The TensorView::data() helper was added to simplify raw-pointer access in CUDA kernels.
• NUM_CHANNELS must include 1 in the build to support single-channel (intensity-only) lidar rendering.