[gsplat] on Windows, spherical harmonics consume excessive VRAM

[machenmusik]

With #2521, and the bicycle dataset (thanks @kerrj), on my Windows 11 machine with CUDA 11.8, ns-train gaussian-splatting colmap --data data\bicycle --colmap-path sparse/0 exceeds 16GB VRAM (this card's limit) after iteration 1810, so iterations are slower starting at 1820, catastrophically slower at 1910, and ultimately fail with CUDA OOM once 32GB VRAM is exceeded, sometime before reaching 2310 steps.

With lower sh-degree values, the pressure is less (although this machine still fails). With sh-degree of zero, training completes successfully without exceeding 16GB VRAM (and thus slowing down).

Per @kerrj , this excessive VRAM consumption does not occur on Ubuntu Linux on similar hardware.

[jb-ye]

This issue might be fixed in the latest PR: https://github.com/nerfstudio-project/nerfstudio/pull/2709

I have successfully trained GS on bicycle dataset.