Clean up test lists to reduce total cost

I was just looking at the model cost of running some CESM tests with MARBL turned on, and it's not good: an `SMS.TL319_t232.G1850MARBL_JRA.derecho_intel` test costs 1000 core-hours (a comparable test without MARBL is in the neighborhood of 80 cpu-hours). I think the bulk of the additional cost comes from having every MARBL diagnostic in the diag_table -- by default MARBL asks MOM6 writes "minimal" output (49 fields) in fully coupled runs, "full" output (238 fields) in ocean-only and FOSI runs, and every diagnostic (353 fields) if the test suite is on.

I think we want to do the following:

- [ ] the `aux_mom_MARBL` test list should continue to write the full output, but tests should be shortened (`SMS_Ld2` instead of `SMS` should cost ~400 cpu-hours, which isn't quite as bad)
- [ ] the `prealpha`, `prebeta`, and `aux_mom` test lists should only either (a) only write the default output based on the compset, or (b) write minimal output for both fully coupled and full output runs
- [ ] @alperaltuntas if we turn on FMS's parallel I/O for the test suite, will `cprnc` still be able to compare new tests to a baseline? I don't want to lose bit-for-bit checks, but if the archiver and test system don't care that each time slice is broken across multiple files that would help reduce cost as well

edit: When I initially posted this, I was looking at cpu-hrs / year rather than total cpu-hrs. I've adjusted the numbers in the opening paragraph, but the point still stands -- we think of MARBL as increasing cost somewhere between 3x and 5x, but the tests are 12x more expensive and I think a lot of the gap between "3-5x" and "12x" is due to I/O