An example of how CMake should form the device linking command line

The existing issue with how CMake does device linking is that it drops all shared libraries and unknown libraries from the command line.

This is problematic as these unknown libraries can have device symbols that you want resolved.

Question 1:

I had originally wondered if the way we pass things to the nvcc -dlink made a difference ('-lcublas_device' versus 'cublas_device.a') but in testing both of these signatures cause the device symbols from cublas_device to be baked into the output object file:

duplicate symbol ___cudaRegisterLinkedBinary_51_tmpxft_0000561c_00000000_15_axpy_compute_60_cpp1_ii_c7deafe8 in:
    /Users/robert/Work/Sandbox/src/CudaDeviceLinking/build/libstatic_1.a(cmake_device_link_1.o)
    /Users/robert/Work/Sandbox/src/CudaDeviceLinking/build/libstatic_2.a(cmake_device_link_2.o)

Question 2:

We still need to do device linking on dynamic libraries, the process still looks to resolve all symbols and will error out if they are not found

Question 3:

Can we at all fix the diamond pattern problem where a static library is embedded in two places?