GraphBLAS/rmm_wrap: Slightly overhaul build rules

[mmuetzel]

Rename target to RMM_wrap and export it with the GraphBLASTargets.
Find package CUDAToolkit and link to CMake target instead of CUDA runtime library name.
Fix error when configured with ENABLE_SHARED_LIBS=OFF.

I can't test this change locally (no working CUDA yet). But I hope it will help for #541.

[DrTimothyAldenDavis]

Thanks! I'll merge it in and test it out, both here and in the draft GraphBLAS 9.0.1 on the GraphBLAS repo.

[DrTimothyAldenDavis]

Works!

I had to make some other fixes, by back-porting those fixes from the draft GraphBLAS 9.0.1. But this gets me back to where I was, in both v8.3.0 and v9.0.1.

I'll post those to v8.3.0 here, even though the CUDA kernels are fully functional yet. But at least now the CUDA and RMM_WRAP libraries both build OK in v8.3.0.

thanks!