Add ROCm-compatible source build fixes

[ZJLi2013]

Summary

This PR makes the existing CUDA source tree buildable with PyTorch ROCm/hipify while keeping the CUDA build path unchanged. It only modifies existing source and packaging files; it does not add generated HIP sources.

Changes

• Skip the CUDA-only --use_fast_math nvcc flag when building with a ROCm PyTorch runtime.
• Remove CUDA-only headers that are unavailable in the ROCm toolchain.
• Normalize CUDA kernel launch syntax so hipify can parse the existing .cu sources.
• Replace CUDA-only trap / pointer type usage with forms accepted by ROCm hipcc.

Reproduction

git clone https://github.com/ZJLi2013/diff-triangle-rasterization.git
cd diff-triangle-rasterization
git checkout rocm_support
git submodule update --init --recursive

export PYTORCH_ROCM_ARCH=gfx942
pip install . --no-build-isolation

python - <<'PY'
import torch
import diff_triangle_rasterization
print(torch.__version__, torch.version.hip)
print("diff_triangle_rasterization import ok")
PY

Tested On

• GPU: AMD Instinct MI300X
• ROCm: 6.4.3
• PyTorch: 2.6.0 ROCm Docker build
• Docker: rocm/pytorch:rocm6.4.3_ubuntu24.04_py3.12_pytorch_release_2.6.0

Notes

• CUDA builds still use the existing cuda_rasterizer/*.cu source path.
• This PR intentionally avoids checking in generated hip_rasterizer sources.