Fix Windows CUDA build: don't pass /openmp:experimental to nvcc

[CaptainArni]

Problem

A CUDA release build on Windows/MSVC (scripts/build_windows.ps1 -Preset windows-cuda-release) fails to compile the two CUDA sources in engine_runtime:

FAILED: CMakeFiles/engine_runtime.dir/src/framework/sampling/torch_random_cuda_runtime.cu.obj
nvcc fatal   : A single input file is required for a non-link phase when an outputfile is specified
FAILED: CMakeFiles/engine_runtime.dir/src/framework/audio/istft_cuda_runtime.cu.obj
nvcc fatal   : A single input file is required for a non-link phase when an outputfile is specified

Reproduced on latest main (365b4f6) with CUDA 13.3 + MSVC (VS 2022/18).

Cause

CMakeLists.txt adds the MSVC OpenMP flag to engine_runtime unconditionally:

target_compile_options(engine_runtime PRIVATE /openmp:experimental)

engine_runtime also compiles CUDA sources (istft_cuda_runtime.cu, torch_random_cuda_runtime.cu), so this flag is applied to their nvcc invocations too. The generated compile line for a .cu object ends with a bare /openmp:experimental:

FLAGS = -Xcompiler=/utf-8 -Xcompiler="-O2 -Ob2" -DNDEBUG -std=c++17 \
        "--generate-code=arch=compute_120a,code=[sm_120a]" -Xcompiler=-MD /openmp:experimental

nvcc does not recognize /openmp:experimental and does not forward it to the host compiler (-forward-unknown-to-host-compiler only forwards --prefixed unknowns, not /-prefixed ones). It therefore parses the token as a second input file, and aborts because an output file is already specified — hence the nvcc fatal above.

Note: passing -DOpenMP_CUDA_FLAGS=/openmp:experimental in build_windows.ps1 does not help, because the offending flag does not come from an OpenMP::OpenMP_CUDA target — it is the hard-coded target_compile_options line, which lands on every language in the target.

Fix

Guard the flag on the C++ language so it never reaches nvcc:

target_compile_options(engine_runtime PRIVATE $<$<COMPILE_LANGUAGE:CXX>:/openmp:experimental>)

The OpenMP 4.0 SIMD pragmas in the C++ sources (e.g. longformer_attention.cpp) still receive the flag; the .cu files no longer do.

Verification

• Before: engine_runtime build fails at the two .cu objects with nvcc fatal (above).
• After: reconfigured and rebuilt engine_runtime on the windows-cuda-release preset — both CUDA objects compile and the target builds to completion (CUDA 13.3, MSVC, RTX 5090 / sm_120a).

[0xShug0]

@CaptainArni Merged. Thank you!