Add ROCm installation notes for CubePart

[ZJLi2013]

Summary

Add ROCm installation and inference notes for CubePart.

Changes

• Document a ROCm PyTorch Docker setup for AMD GPUs.
• Add requirements_rocm.txt with helper build dependencies and NumPy/scikit-image pins.
• Note that example meshes require Git LFS assets.
• Document the current warp-lang geometry fallback caveat on ROCm.

Tested on

• GPU: AMD Instinct MI300X / gfx942
• ROCm: 6.4.3
• PyTorch Docker: rocm/pytorch:rocm6.4.3_ubuntu24.04_py3.12_pytorch_release_2.6.0

Results

• pip install -r requirements_rocm.txt && pip install -e . --no-build-isolation succeeded.
• examples/run_inference.py completed on examples/inputs/jellyfish_car.glb.
• A 50-step demo run exported 8 part meshes plus a combined parts.glb.

Notes

• The ROCm instructions keep the container-provided ROCm torch and torchvision wheels.
• warp-lang may fall back during geometry extraction, but mesh export still completes.