Add build configured torchfort_config.h header to indicate GPU support in build.

[romerojosh]

Our current torchfort.h and torchfort_rl.h header files use #ifdef ENABLE_GPU preprocessor macros to control whether cudaStream_t is redefined to enable CPU only builds. The issue is that this macro is only defined when building the library, so users building external applications with CUDA can run into problems when including these headers. In particular, if they don't also define ENABLE_GPU in their compilation, they will run into redefinition errors.

For example, compiling the following program:

#include <torchfort.h>
#include <cuda_runtime.h>

int main(int argc, char* argv[]) {
}

using a typical compilation line like:

mpicxx -I${TORCHFORT_ROOT}/include -I${CUDA_ROOT}/include example.cpp

results in the following error:

"/opt/nvidia/hpc_sdk/Linux_x86_64/25.7/cuda/include/driver_types.h", line 3035: error: invalid redeclaration of type name "cudaStream_t" (declared at line 23 of "/opt/torchfort/include/torchfort.h")
  typedef __device_builtin__ struct CUstream_st *cudaStream_t;

which can be resolved by adding -DENABLE_GPU to the compile line, which is not great user experience.

To fix this, this PR introduces a build time created torch_config.h which will define TORCHFORT_ENABLE_GPU to indicate whether TorchFort was built with GPU support or not. This configuration header is then included in torchfort.h and torchfort_rl.h to control whether cudaStream_t is redefined or not, removing the need for users to define ENABLE_GPU themselves anymore.

[romerojosh]

/build_and_test

[github-actions]

🚀 Build workflow triggered! View run

[github-actions]

✅ Build workflow passed! View run

[azrael417]

this is elegant, thanks for fixing this.