fix: add GPU Docker runtime support

[Artavazd Balaian (REASY)]

Summary

Adds a Docker/Compose runtime for GPU-backed TRELLIS.2 inference.

• Adds a CUDA 12.4 image with PyTorch 2.6 and the required CUDA extensions.
• Adds a Compose service with NVIDIA GPU access, Gradio port forwarding, persistent Hugging Face cache, and output mounts.
• Documents Docker usage, Hugging Face token forwarding, and DINOv3 gated-model access.
• Pins compatible OpenCV and Pillow releases for EXR environment-map loading and Gradio WebP support.
• Updates DINOv3 feature extraction to support both legacy and Transformers 5 encoder layouts.
• Keeps generated models and outputs out of Git.

Validation

• Built the Docker image successfully.
• Verified CUDA access from the container on an RTX 4090.
• Verified FlashAttention, nvdiffrast, and O-Voxel imports.
• Verified OpenEXR decoding of the bundled HDR environment map.
• Verified DINOv3 feature extraction on cuda:0.

[Artavazd Balaian (REASY)]

@microsoft-github-policy-service agree