[ROCm] Add AMD GPU support for TRELLIS.2

[andyluo7]

Summary

Enable TRELLIS.2 to run on AMD Instinct GPUs (MI300X / gfx942) with ROCm.

Your setup.sh already detects ROCm and installs ROCm PyTorch + ROCm flash-attention, which is great! However, the C++ extensions (O-Voxel) and rendering libraries (nvdiffrast) don't build/work on ROCm. This PR fills those gaps.

Changes

O-Voxel HIP port (5 files)

• Add #ifdef __HIP_PLATFORM_AMD__ guards for CUDA→HIP header mapping in all .cu files
• setup.py: default arch gfx942 for AMD MI300X

nvdiffrast ROCm adapter (2 new files)

• trellis2/renderers/nvdiffrast_rocm_adapter.py: Pure PyTorch drop-in replacements for dr.rasterize(), dr.interpolate(), dr.texture(), dr.antialias(), dr.DepthPeeler
• trellis2/renderers/rocm_compat.py: Auto-patches import nvdiffrast.torch as dr when nvdiffrast is unavailable

Companion PR

• CuMesh HIP port: [ROCm] Add HIP support for AMD Instinct GPUs JeffreyXiang/CuMesh#31 (15 files, same approach)

Testing

• ✅ AMD MI300X (gfx942), ROCm 7.0.2, PyTorch 2.9.1
• ✅ O-Voxel compiles with hipcc
• ✅ CuMesh compiles with hipcc (import cumesh works)
• ✅ FlexGEMM compiles on ROCm (no changes needed)
• ✅ Pipeline class imports and loads pretrained model
• ✅ Rasterize/interpolate/texture adapter verified with unit tests
• ✅ All code cross-compilable — CUDA builds unaffected

Usage on ROCm

# Add this before importing renderers:
import trellis2.renderers.rocm_compat

# Then use normally
from trellis2.pipelines import Trellis2ImageTo3DPipeline
pipeline = Trellis2ImageTo3DPipeline.from_pretrained('microsoft/TRELLIS.2-4B')
pipeline.cuda()
result = pipeline.run(image)

Known Limitations

• nvdiffrast rendering uses pure PyTorch fallback (no antialiasing, slower than CUDA rasterizer)
• nvdiffrec PBR lighting is stubbed (not ported)
• flash-attention falls back to SDPA on ROCm 7.0.2 (works but slower)
• Only tested on gfx942 (MI300X)