fix(splatfacto): replace hardcoded .cuda() with .to(self.device) for MPS support

Two lines in splatfacto.py use .cuda() directly instead of the model's
device property, causing crashes on non-CUDA systems (Apple Silicon MPS,
CPU-only machines) with "Torch not compiled with CUDA enabled".

Replace with .to(self.device) which correctly resolves to cuda:N on NVIDIA,
mps:0 on Apple Silicon, or cpu as appropriate.

Tested on M2 Max (MPS) — model initializes and runs forward pass successfully.
CUDA compatibility preserved: .to(self.device) resolves to cuda:N on CUDA systems.

Discovered and benchmarked by an autonomous Claude Code agent ("Ralph")
optimizing a 3D scanning pipeline for Apple Silicon.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: ChrisMOxon

nerfstudio/models/splatfacto.py

@@ -205,7 +205,7 @@ def populate_modules(self):
             # We can have colors without points.
             and self.seed_points[1].shape[0] > 0
         ):
-            shs = torch.zeros((self.seed_points[1].shape[0], dim_sh, 3)).float().cuda()
+            shs = torch.zeros((self.seed_points[1].shape[0], dim_sh, 3)).float().to(self.device)
             if self.config.sh_degree > 0:
                 shs[:, 0, :3] = RGB2SH(self.seed_points[1] / 255)
                 shs[:, 1:, 3:] = 0.0
@@ -532,7 +532,7 @@ def get_outputs(self, camera: Cameras) -> Dict[str, Union[torch.Tensor, List]]:
         camera_scale_fac = self._get_downscale_factor()
         camera.rescale_output_resolution(1 / camera_scale_fac)
         viewmat = get_viewmat(optimized_camera_to_world)
-        K = camera.get_intrinsics_matrices().cuda()
+        K = camera.get_intrinsics_matrices().to(self.device)
         W, H = int(camera.width.item()), int(camera.height.item())
         self.last_size = (H, W)
         camera.rescale_output_resolution(camera_scale_fac)  # type: ignore