Add cuda_device_arch

[michael-kenzel]

This little hack allows us to expose the CUDA target architecture as a device function. It exploits the fact that all relevant CUDA C++ compilers need to work with the nvvm libdevice library and, thus, implement transformations around __nvvm_reflect which we can use to get at the target architecture.

tested and seems to work with NVCC, NVRTC, and Clang

[michael-kenzel]

found better approach -> closed