feat(docker): Add nvdiffrast to Docker image for faster texturing

[hummat]

Summary

nvdiffrast is used by sdfstudio's sdf-texture-mesh for GPU-accelerated UV rasterization. Currently it's not installed in the Docker image, so texturing falls back to slower CPU rasterization.

Motivation

• GPU rasterization is significantly faster for texture baking
• Enables future integration with nvdiffrec for material refinement (feat(export): PBR-ish texture extraction and BRDF-aware training #11 Stage 3)

Context

In sdfstudio/exporter/texture_utils_v2.py:

• nvdiffrast is optionally imported with NVDIFFRAST_AVAILABLE flag
• Falls back to rasterize_uv_cpu() when unavailable
• GPU path uses dr.RasterizeCudaContext() for faster rasterization

Proposed Solution: Builder Wheel

Follow the same pattern mini-mesh uses for tiny-cuda-nn.

Important: Use --no-build-isolation when pip installing CUDA extensions to ensure correct PyTorch ABI compatibility (nvdiffrast docs).

# ===== Builder stage =====
# nvdiffrast wheel
RUN cd /workspace/git && git clone https://github.com/NVlabs/nvdiffrast.git && cd nvdiffrast && \
    TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST}" pip wheel . --no-build-isolation --no-deps -w /workspace && \
    cd /workspace/git && rm -rf nvdiffrast

# ===== Runtime stage =====
COPY --from=builder /workspace/nvdiffrast*.whl /tmp/
RUN pip install --no-cache-dir /tmp/nvdiffrast*.whl && rm /tmp/nvdiffrast*.whl

Alternatives Considered

1. Runtime install: pip install --no-build-isolation git+https://github.com/NVlabs/nvdiffrast.git — may work if PyTorch's bundled CUDA is sufficient, but untested in Docker runtime
2. Skip: Keep CPU fallback, accept slower texturing

Tasks

• Add nvdiffrast wheel build to Docker builder stage
• Install wheel in runtime stage
• Test GPU rasterization works in container
• Update Docker build docs

References

• https://github.com/NVlabs/nvdiffrast
• https://nvlabs.github.io/nvdiffrast/ (Docker build notes)
• sdfstudio/exporter/texture_utils_v2.py:46-50 (import)
• sdfstudio/exporter/texture_utils_v2.py:482-485 (fallback logic)
• mini-mesh tiny-cuda-nn pattern: docker/Dockerfile:109-115, 176-177

Related Issues

• feat(docker): Add gsplat to Docker image to avoid runtime compilation #8 — gsplat (similar Docker build pattern)
• feat(export): PBR-ish texture extraction and BRDF-aware training #11 — PBR texture extraction (Stage 3 depends on nvdiffrast for nvdiffrec)