- Bumped cmake minimim version to 3.17
- Changed `DACE_LIBS` element to `CUDA::cudart`
- Removed unneeded `include_directories` and link_directories calls
- Removed `compile_cuda`. CUDA files are now passed directly to targets
- Removed `-fPIC` and `-std` args from nvcc as they are handled
  automatically now
- Renamed `CUDA_NVCC_FLAGS` to `CMAKE_CUDA_FLAGS`

- Moved `-gencode` handling to `cmake`: cmake variable
  `DACE_CUDA_ARCHITECTURES_DEFAULT` is set in python code to be used by
  `CMAKE_CUDA_ARCHITECTURES` instead of manually creating the compiler arg string.
- Default cuda arch is no longer included forcefully in presence of a native architecture
- `get_cuda_arch.cpp` now returns a properly formatted architecture
  string compatible with cmake.
- `get_cuda_arch.cpp` now fails if no architectures are found

This reverts commit abdfc88.