Alpaka: Make sure that the kernel function returns void with `ALPAKA_ACC_GPU_CUDA_ONLY_MODE`

[BenjaminW3]

We should check at compile time that the TKernelFnObj passed into the executor returns void.
There is no way to handle functions returning a value because each thread produces one.

[BenjaminW3]

With #695 an static_assert has been added that checks the return type.
It is disabled when ALPAKA_ACC_GPU_CUDA_ONLY_MODE is set. This restirction should be investigated and removed.

[ax3l]

Seems to be a problem with pure __device__ functions only: https://github.com/ComputationalRadiationPhysics/alpaka/pull/695#issuecomment-446103194

[psychocoderHPC]

But as I can see from the comments in #695 we can get the correct return type on the device side. This means we can add a check for CUDA function within the device code.
I will open a PR for it.

[ax3l]

I don't see yet how since the signature does not contain void but just some generic int... surprise me :)

[psychocoderHPC]

I opened #711 with the implementation for the device side validation for CUDA and HIP (to guard the case that CUDA is used as HIP backend)