[ROCm] Add AMD GPU support via PyTorch hipify

[jeffdaily]

This adds AMD GPU (ROCm) support. The CUDA backend is a PyTorch CUDAExtension, so building it against a ROCm build of PyTorch runs hipify at build time to translate the CUDA sources to HIP automatically. No changes to the CUDA kernel or host C++ sources are required, and the CUDA build path is unchanged.

Two changes support the ROCm build:

• setup.py: on Windows, c10.dll does not export the inherited c10::ValueError(SourceLocation, string) constructor (MSVC does not re-export inherited constructors even for C10_API classes). Headers pulled in via <torch/extension.h> trigger TORCH_CHECK_VALUE, which generates a dllimport reference to that constructor and fails to link with LNK2001. A win32-only /ALTERNATENAME linker directive redirects the missing thunk to c10::Error(SourceLocation, string), which is exported; ValueError IS-A Error with no additional data members, so the redirect is semantically safe. The directive is guarded by sys.platform == "win32" and does not affect Linux or the CUDA build. This is fixed upstream by [c10] Fix missing symbol exports for ValueError/NotImplementedError on Windows pytorch/pytorch#175340, so the directive can be removed once that fix is in the minimum supported PyTorch.
• README: a note in the install section explaining that the ordinary pip install works on ROCm and that PYTORCH_ROCM_ARCH selects the target architecture.

Test Plan

Build against a ROCm PyTorch and run the suite on an AMD GPU:

PYTORCH_ROCM_ARCH=gfx1100 pip install -e .
pytest tests/

On Linux gfx1100 (Radeon Pro W7800, ROCm 7.2.1) and Windows gfx1201 (RX 9070 XT) the full suite passes 3/3 (test_simple, test_cuda_equal_to_cpu, test_compare_to_scipy). On Linux gfx90a (MI250X) the assignment checks pass with valid, optimal, deterministic GPU assignments. The CUDA build path is unchanged.

Authored with assistance from Claude.