Adds ROCm/HIP support for the CUDA GPU backend from #25, so fpie can build the GPU solver for AMD GPUs while keeping the existing NVIDIA/CUDA path unchanged. The release also documents ROCm/HIP build usage, auto-enables HIP during source installs when a ROCm toolchain is present and nvcc is absent, and includes lint/spelling cleanup from the follow-up CI fix.
PyPI: https://pypi.org/project/fpie/0.3.4/
Full Changelog: v0.3.3...v0.3.4