Enable building and testing Omega in single precision

[mwarusz]

This PR enables configuring Omega to use single precision and adds a test (reusing the tendency terms test) checking that we can build and run in single precision. The main changes are:

• Changing R8 to Real where appropriate
• Reading mesh and state arrays into temporary R8 arrays before storing them into the class members
• Always using R8 buffers for halo exchanges
• CMake logic to build the single precision library when needed

The halo buffers change is not optimal for performance in single precision. It would need to be optimized if we decide
to pursue this option seriously in the future.

Checklist

• Documentation:
  • User's Guide has been updated
  • Developer's Guide has been updated
  • Documentation has been built locally and changes look as expected
• Testing
  • CTest unit tests for new features have been added per the approved design.
  • Unit tests have passed. Please provide a relevant CDash build entry for verification.

[sbrus89]

This looks good to me and includes some nice clean-ups as well. I just had a couple questions related to the MPI calls and _Real suffix.

[sbrus89]

These function calls are a nice improvement, thanks.

[grnydawn]

Both the double-precision and single-precision versions of the Omega unit tests were profiled using Nsight profilers, and the results indicate that this PR correctly generates the corresponding GPU kernels. However, I have not verified the validity of the algorithms.

The profiling results indicate that the unit tests in this PR are insufficient to determine whether single-precision improves performance, as the kernel sizes are too small and do not appear to be representative of typical climate algorithms.

I approve this PR, assuming it properly handles merging commits from external libraries as I noted in another comment.

[grnydawn]

I have profiled PLANE_TEST unit test with Nsight profilers and summarized the profiling result below. Further details are in (https://acme-climate.atlassian.net/l/cp/YSFk1Zra).

• The PR correctly generates the single-precision version of the Omega TEND_PLANE unit test.
• The elapsed time for both the single-precision and double-precision versions of the unit test is nearly the same, at around 155 ms, excluding initialization and finalization routines.
• GPU resources are underutilized in both cases. For the double-precision version, Compute (SM) Throughput is 4.15%, and Memory Throughput is 1.39%. For single-precision, Compute (SM) Throughput is 7.27%, and Memory Throughput is 0.81%.
• It appears that the kernels are too small to effectively compare the performance characteristics of different floating-point precisions. The longest kernel runs for about 27 µs, but most kernels run in under 20 µs.
• The arithmetic intensity (AI) of the kernels also appears to be too high, meaning these kernels might not accurately represent the performance characteristics of the full Omega model. The typical AI of climate algorithms ranges between 0.1 and 1, but the AI of these unit test kernels reaches up to 460 FLOPs/byte.

[brian-oneill]

After the rebase, able to build and run tests successfully on Chrysalis and Perlmutter CPU & GPU. Everything looks good, approving.

[sbrus89]

Approving based my inspection and testing by @brian-oneill and @grnydawn.