Make alpaka follow CMAKE_CUDA_RUNTIME_LIBRARY

[fwyzard]

CMake uses CMAKE_CUDA_RUNTIME_LIBRARY to link the static (libcudart_static.a) or shared (libcudart.so) CUDA runtime library.

This PR makes alpaka follow the same variable, and applies it also to the cuRAND libraries.

[psychocoderHPC]

By default CMAKE_CUDA_RUNTIME_LIBRARY is not set so this if condation will be false and we link static.
In all old alpaka versions we linked always against cudart.
@fwyzard was the change of the default behavior of alpaka intended?

I run into issue with our main code PIConGPU when compiling with clang as CUDA compiler.

If there is no strong reason for the change of the default behaviour I could open a PR with

if(NOT DEFINED CMAKE_CUDA_RUNTIME_LIBRARY) 
     set(CMAKE_CUDA_RUNTIME_LIBRARY "Shared") 
 endif() 

Which will keep the default behaviour of alpaka but still provide the possibility of static linking.

[fwyzard]

No, I did not mean to change the default behaviour.

However, the default behaviour before #2329 and this PR was kind of tricky.

If configured with alpaka_DISABLE_VENDOR_RNG=OFF:

• the CUDA binaries (e.g. axpyTest) are linked with both libcudart.so and -lcudart_static; I think the shared library took precedence because it was coming first on the command line;
• the CUDA host-only applications (e.g. hostOnlyAPITest) were linked with libcudart.so.

If configured with alpaka_DISABLE_VENDOR_RNG=ON:

• the CUDA binaries (e.g. axpyTest) were linked with cudart_static;
• the CUDA host-only applications (e.g. hostOnlyAPITest) were not linked with either library, and in fact failed to link.

[fwyzard]

If the default should be "shared" (I agree it should be), let me invert the logic of the checks:

        if(CMAKE_CUDA_RUNTIME_LIBRARY STREQUAL "Static")
            target_link_libraries(alpaka INTERFACE CUDA::cudart_static)
        else()
            target_link_libraries(alpaka INTERFACE CUDA::cudart)
        endif()

[fwyzard]

See #2348 .