fix restart issues that arise with time averaged fields

[mvertens]

Summary

Restores bit-for-bit reproducibility across restarts by reverting the
esm_history_vars default in cime_config/namelist_definition_cism.xml
from _tavg variants back to instantaneous fields for both gris and
ais ice sheets.

Root cause

In cismwrap v8 the default esm_history_vars list was changed from
instantaneous fields (e.g., acab_applied, calving_rate, total_smb_flux)
to their time-averaged variants (acab_applied_tavg, calving_rate_tavg,
total_smb_flux_tavg, etc.). With these _tavg variants in the history
list, the model no longer reproduces bit-for-bit across a restart.

The _tavg fields are auto-generated by utils/build/generate_ncvars.py
in CISM. The codegen explicitly sets load: 0 on every generated variant,
so the _tavg accumulator arrays are never written to or read from the
CISM restart file. The per-output-file scalar (outfile%total_time) that
normalizes the accumulator at write time is also reset whenever a new
output file is opened post-restart. As a result, a continuous run and a
restart run accumulate over different windows and divide by different
total_time values, producing different history-file values at the same
wall-clock time.

The prognostic state is still bit-for-bit across restart — only the
diagnostic _tavg outputs differ — but any BFB check that compares
history files flags the run as a failure.

This was masked in earlier wrapper releases because esm_history_vars
listed only instantaneous fields, which are deterministic at the output
moment and do not depend on accumulator state.

What this PR changes

• cime_config/namelist_definition_cism.xml: for the two
  cism_evolve_this_icesheet=".true." entries (gris and ais), drop
  the _tavg suffix from all fields. frontal_melt_rate_tavg is mapped
  to the corresponding instantaneous variable melt_rate. All other v8
  additions (per-icesheet split, the extra prognostic fields like
  velnorm, btemp, tempstag, powerlaw_c, base) are kept.
• Added an inline comment explaining the reason for the revert.

Trade-off

The history files lose time-averaging in favour of snapshot values at the
output moment. Time-averaged diagnostics can be reconstructed post-hoc, or
reintroduced after the upstream CISM _tavg restart-save issue is fixed.

Upstream issue

The proper fix lives in CISM: both the <var>_tavg arrays
(generate_ncvars.py line 126) and the outfile%total_time scalar
(libglimmer/ncdf_template.F90.in) need to be saved in the restart file
together. That work is tracked separately.

Test plan

• Rebuild namelists (./preview_namelists) on an existing case and
  confirm esm_history_vars in cism.<icesheet>.config no longer
  contains any _tavg entries.
• Run the 3-year continuous + restart-at-year-2 test and verify the
  year-3 history files are bit-for-bit identical between the continuous
  and restart runs.
• Confirm prognostic answers are unchanged relative to a pre-revert
  run that didn't trigger the BFB-restart check (i.e., the model
  trajectory itself is unaffected — only what gets written to history).