update pgi compiler optimization to address reproducibility problems

[apcraig]

Update pgi compiler optimization, switch from O2 to O.

• Developer(s): tcraig

• Are the code changes bit for bit, different at roundoff level, or more substantial? NOT bit-for-bit for pgi compiler, but this is good.

• Is the documentation being updated with this PR? (Y/N) N
  If not, does the documentation need to be updated separately at a later time? (Y/N) N

• Other Relevant Details:

Confirmed that O0, O1, and O all generate results that are reproducible as expected (different block sizes, etc). O2 does not. O is between O1 and O2 and seems to actually perform better than O2 on conrad and gordon.

[apcraig]

@eclare108213, I would like to get this onto the master in the next day so it's part of the weekend testing on conrad and gordon. This will establish a full set of new baselines for the pgi compiler to make it easier to test other change (like dynamic allocation). This will change answers which I believe is a good thing.

Feel free to review and Pull, this is ready and has been tested, hash 9a08fd8 on https://github.com/CICE-Consortium/Test-Results/wiki/cice_by_hash_forks

[eclare108213]

Works for me. If someone else runs into pgi compiler problems, I'm sure we'll hear about it...

[eclare108213]

Actually, do you want -g even when the ICE_BLDDEBUG flag is false?

[apcraig]

The -g doesn't hurt for these runs. I would leave -g for now. pgi is so unreliable as it is.

Also, I have change the pgi settings for all our pgi machines, conrad and gordon. Those are the only machines we test pgi on right now.