GitHub: FMS-MOM6 build jobs, tarball artifacts

[marshallward]

Inconsistencies in jobs within a workflow were leading to differences in compiler versions across jobs. This was breaking GitHub Actions workflows.

We want to continue using *-latest versioning, so this patch merges the FMS build into each MOM6 build. Although there is greater compute time, it should be happening in parallel. It may even be shorter, since we do not need to wait on the additional OS setup.

This patch also packs executables and other data into tarballs. This preserves most of the file data (file mode permissions and modified-time, mainly) so that we don't have to do Makefile tricks or use chmod to get things working in the next job. We also don't need to resort to horrible awk tricks to generate executable lists.

There may be a slight speedup here if the builds happen in parallel, since we are doing one less image setup, but CI runtimes are very variable so I don't think we'll be able to know right away.

(@Hallberg-NOAA noted that different MOM6 builds may still have different versions. Eventually we may need to have one job set the version and then force the others to use this prescribed version. But this is an improvement over the current state.)

[marshallward]

There is some issue here where FMS is being rebuilt in the test jobs, even though the executables are already built. (Also a minor typo in build-opt-target and a problem building the tc4 grid data in test-regressions.)

[marshallward]

I have many multiple adjustments to this PR:

• The "oldest" -o make flags used to prevent executable rebuilds have been restored where needed

• Some missing DO_REGRESSION_TESTS flags have been restored (also fixing some redundant builds)

• Perf profiling output has been repaired

• compare.timing output has been repaired

• Several rules which must be run serially now have .NOTPARALLEL directives

• the URL-tracking macros (MOM_TARGET_SLUG, MOM_TARGET_LOCAL_BRANCH) are now defined at the workspace environment variable level

• The unused testing-setup macro has been deleted

• tarball unpacking is now verbose, for output inspection

• .gcno files are now "touched" to disable modified-time warnings

• Dimension tests are now packed as a single matrix of tests

• the TARGET_CODEBASE checkout rule is now wrapped in a DO_REGRESSION_TESTS, to avoid an empty : rule.

  However the whole DO_REGRESSION_TESTS usage has shown itself to be very problematic and did not properly consider the "parsing" vs "execution" stages of make. This needs a tidy up, so that we can remove the DO_REGRESSION_TESTS flags from the yaml file.

While there are even more improvements to be made, it's probably best that I stop while I am ahead.