Need coverage for building OMP/TBB algorithms with nvcc

This is a hole in our test coverage that's being used in the wild, leading to issues such as NVIDIA/cccl#743.

We currently build all DEVICE=CUDA configurations with nvcc as CUDA CMake targets, but all other configurations are built directly with the host compiler as CXX CMake targets. We have no coverage for non-CUDA device algorithms built through nvcc.