#356 HIP: build the AMD engine on Windows

本文

Windows can compile the HIP engine against an installed ROCm 10 or HIP SDK.

- MSVC host compiles force-include `cuda_runtime.h` with `/FI` and forward slashes. HIP compiles still use `-include`.
- `setup.py` finds that ROCm tree (`ROCM_PATH`, `rocm-sdk`, `HIP_PATH`, or the HIP SDK) and uses its clang for the host as well as for HIP. CMake will not mix `cl.exe` with Clang HIP.
- On Windows only, a successful gfx1201 hipBLAS GEMM may leave `hipErrorInvalidValue` set. The engine clears that sticky error so the next kernel check does not exit. Linux still reports it.

This is the Windows build slice of #247, rebased onto 0.1.31. It does not include the README rewrite, the multi-shard pack, the iGPU ordinal fix, or the copy batching.