For pm-cpu/pm-gpu, update some module versions to current machine defaults

[ndkeen]

Minor version increases for several modules.
Does not impact compiler versions.
Motivation is to keep up-to-date with machine defaults.
Do not see any measurable performance changes.

cray-mpich/8.1.22 -> cray-mpich/8.1.24
cray-hdf5-parallel/1.12.2.1 -> cray-hdf5-parallel/1.12.2.3
cray-netcdf-hdf5parallel/4.9.0.1 -> cray-netcdf-hdf5parallel/4.9.0.3
cray-parallel-netcdf/1.12.3.1 -> cray-parallel-netcdf/1.12.3.3
cmake/3.22.0 -> cmake/3.24.3

Added specific version numbers for craype and cray-libsci to reduce surprises when the default version is changed (these were already using default)

Also added a couple of modules to remove just in case they are loaded.

Updating alvarez the same way, but it may be that the machine goes away.

Fixes #5525
[bfb]

[ndkeen]

I tested against baselines on pm-cpu with e3sm_integration. And have been running larger cases with these modules versions in a scream repo (to test pm-gpu).

Example of some of the changes:

  Changes in Cray MPICH 8.1.24

      - CAST-24802 - Fix for MPI_CXX datatypes in the mpi header
      - CAST-26727 - Error instead of warning when NIC asymmetry is detected at start-up
      - CAST-31527 - Remove shared object constructors in MPI-IO
      - PE-44058 - Remove fallback usage of VNI and depend on launcher definitions
      - PE-44653 - Fix regression with memory debugging in OFI code
      - PE-44772 - Add support for collecting CXI counters
      - PE-44989 - Fix MPI-IO debug trace output
      - PE-45030 - Fix environment variable printing for MPI_DPM_DIR and MPICH_SPAWN_USE_RANKPOOL
      - PE-45042 - Add support for lmod auto swapping to the Cray-MPICH module
      - PE-45094 - Add support for program environment swapping and Cray-MPIXlate
      - PE-45160 - Enable GPU kernel-based optimizations by default

https://github.com/Parallel-NetCDF/Parallel-NetCDF.github.io/blob/master/Release_notes/1.12.3.md

[ndkeen]

There are some TPUT failures when I tried to run again vs baselines, but it looks fine to me -- just comparing two very fast cases. Would I need to bless TPUT fails? Or increase the TPU tolerance for the machine?

[ndkeen]

Merged to next. The tests runs are way too short to actually see any performance differences between before/after. In my testing, some tests report larger than 10% TPUT, but those cases seem fine and I have not seen any issues with larger/longer cases.

Actually, I forgot that NERSC current has an env variable set (FI_MR_CUDA_CACHE_MONITOR_ENABLED=0) to avoid NODE_FAILs, but is slowing many things down. Which could be the reason for some of the TPUT errors.

Looks like all the GNU tests passed on cdash. Manually checking nvidia developer -- completed as expected.