CUDA separable compilation broken

Additional information after investigation

The issue first appears after this PR was merged: https://github.com/LLNL/RAJA/pull/1004 (10/19/2022).
Code compiles, but all CUDA tests error at run time and report invalid device function at https://github.com/LLNL/RAJA/blob/develop/include/RAJA/policy/cuda/MemUtils_CUDA.hpp#L229
Verbose build output is identical before and after the merge.
@rhornung67 has been reproducing by comparing builds after the PR merge listed above and the previous PR merge: https://github.com/LLNL/RAJA/pull/1273 (10/18/2022). Specifically (but not sure how much of this matters),
- Edit the top-level CMakeLists.txt file to set min req cmake version to 3.20
- Edit ./scripts/lc-builds/blueos_nvcc_clang.sh
  - load cmake/3.20.2
  - Add lines: -DCMAKE_CUDA_ARCHITECTURES=70 \ -DCMAKE_CUDA_SEPARABLE_COMPILATION=On \
    - Generate build space: ./scripts/lc-builds/blueos_nvcc_clang.sh 11.2.0 sm_70 ibm-14.0.5 (based on initial issue report)

LLNL / RAJA