Add support for NVIDIA Blackwell GPUs (sm_120, RTX 50-series)

- Bump gsplat to >=1.5.3 (sm_120-capable release line) and nerfacc to >=0.5.3
- Drop EOL Python 3.8/3.9: requires-python >=3.10, CI publish/doc workflows
- torch.load(..., weights_only=False) for torch>=2.6 checkpoint compatibility
- Use torch.amp custom_fwd/custom_bwd with device_type="cuda" (torch>=2.4 API)
- Replace deprecated PIL private encoder API in pil_to_numpy with np.asarray
  (fixes Pillow>=11 compatibility)
- Docs: add Torch 2.8+ / CUDA 12.8 install tab for Blackwell GPUs

Tested on RTX 5060 (sm_120) with torch 2.11.0+cu128: splatfacto and
nerfacto train end-to-end and ns-render produces correct output.

.github/workflows/doc.yml

@@ -15,7 +15,7 @@ jobs:
       - uses: actions/checkout@v3
       - uses: actions/setup-python@v3
         with:
-          python-version: '3.9'
+          python-version: '3.10'
       - name: Install dependencies
         run: |
           pip install uv

.github/workflows/publish.yml

@@ -16,7 +16,7 @@ jobs:
       - name: Set up Python
         uses: actions/setup-python@v1
         with:
-          python-version: '3.8'
+          python-version: '3.10'
       - name: Install dependencies
         run: |
           python -m pip install build twine