gsplat CUDA inference adaptation for Gobot

Upstream project: https://github.com/nerfstudio-project/gsplat
Upstream commit: bf302ef532a25175e44bd6bf7f777f599570748c
Upstream commit date: 2026-07-20

The projection, spherical-harmonics evaluation, tile binning, depth sorting,
and front-to-back Gaussian compositing implemented here are derived from the
Apache-2.0 gsplat CUDA rendering and inference sources. This adaptation is not
the upstream ATen-based HiGS macro-tile implementation.

Gobot modifications:
- removed ATen, PyTorch, pybind, training, autograd, and runtime JIT dependencies;
- replaced tensor ownership with an opaque raw-CUDA renderer and persistent buffers;
- added external CUDA stream support for LuisaCompute context integration;
- added packed RGBA8 output and optional geometry-coverage composition;
- limited the implementation to forward inference for one pinhole camera.
