Add ROCm/HIP support (or document the silent CPU fallback on AMD GPUs)

[Bharathbabu-01]

Summary

On an AMD GPU with ROCm, torchmcubes builds CPU-only without any error, then
falls back at runtime. Because PyTorch's ROCm build makes torch.cuda.is_available()
return True (HIP maps onto the CUDA API surface), downstream libraries assume
the GPU kernel path exists. The result is a silent performance cliff rather than
a failure.

Environment

• GPU: AMD Radeon Graphics, gfx1100 (Navi 31), 48 GB
• ROCm 7.2.1, PyTorch 2.9.1 (ROCm build), Python 3.12, Ubuntu 24.04
• torch.cuda.is_available() → True
• torch.cuda.get_device_name(0) → AMD Radeon Graphics

Build output

-- Looking for a CUDA compiler
-- Looking for a CUDA compiler - NOTFOUND
CMake Warning at CMakeLists.txt:25 (message):
NO CUDA INSTALLATION FOUND, TRYING TO INSTALL CPU VERSION ONLY!

The build succeeds. At runtime:
Torchmcubes was not compiled with CUDA support, use CPU version instead.

Why this is worth flagging

ROCm systems have no nvcc, so the CUDA compiler probe can never succeed, but
they do have hipcc and a fully capable GPU. The current CMake logic treats
"no nvcc" as "no GPU".

Measured impact in TripoSR on the hardware above, mean of 14 assets:

Stage	Time	Device
Transformer inference	1051 ms	GPU
Marching cubes (this library)	1332 ms	CPU fallback

Surface extraction now costs more than the model itself, and is the single
largest non-GPU stage in the pipeline.

Suggested directions

1. HIP build path. PyTorch ships torch.utils.hipify, which mechanically
   converts CUDA sources to HIP. The kernels here look like good candidates
   since they use standard CUDA constructs rather than vendor-specific intrinsics.
2. Detect HIP in CMake. Probe for hipcc / HIP_PATH alongside the CUDA
   check so ROCm systems take a GPU path instead of falling through.
3. At minimum, document it. A README note that ROCm is unsupported and will
   silently build CPU-only would save people from diagnosing a slowdown that
   looks like a model problem.

Offer

I hit this while building an image-to-3D pipeline on Radeon hardware and worked
around it downstream. Happy to test a HIP build on gfx1100 and report back, or
to help with the CMake detection change if that's useful.