Enable C++ scream P3 and SHOC on CPUs, Nvidia GPUs, and AMD GPUs

[mrnorman]

This PR enables C++ scream P3 and SHOC on the CPU within PAM.

• User can build P3 and SHOC from an existing scream clone entirely within PAM in standalone fashion
• User then builds the Kokkos-facing interface layers for P3, SHOC, deallocating globals, and finalizing kokkos. The header files for these are library agnostic (no kokkos or YAKL). Only the source files have Kokkos included.
• Kokkos is initialized as needed in p3 and shoc C++ routines if it isn't already initialized
• P3 tables are allocated and read as needed inside p3 C++ routine and stored in globals so they persist. They are only allocated and read upon first call to P3, and the rest of the MMF code doesn't need to know about it.
• The pam::deallocate_scream_cxx_globals() routine deallocates all C++ scream globals, which are currently only the p3 tables
• The pam::call_kokkos_finalize() routine finalizes Kokkos at the end of the standalone model run.
• A workflow for running standalone is documented here: https://github.com/E3SM-Project/PAM/blob/mrnorman/cxx-scream-physics/standalone/mmf_simplified/README.md
• Kokkos and YAKL are never used in the same header file, source file, or compilation unit. They are kept in entirely separate libraries.
• The Fortran-based P3 and SHOC are still in the code, and they give the same answer on the CPU as C++ does for now.
• Nvidia GPU code runs and gives physically realistic results.
• AMD GPU code runs and gives physically realistic results.
• When using the C++ P3 and SHOC routines, the user must now provide the number of vertical levels in the PAM/standalone/machines/lnux* environment files to be used in the cmakescript.sh scripts when building. Scream C++ code needs this knowledge at compile time.
  • A check is implemented in the standalone driver to ensure the YAML file's nz matches the nlev Scream is compiled with.
  • This check is only if C++ scream routines are used. Fortran scream can still let nlev / nz be a runtime variable.

[mrnorman]

Update: PAM now compiles C++ scream physics and interfaces successfully on Nvidia GPUs including Summit.

[mrnorman]

We can now run PAM successfully on CPUs, Nvidia GPUs, and AMD GPUs end-to-end on the device using C++ scream P3 and SHOC. Big thanks to @bartgol for help with the scream compilation side.