enable loop parallelism in NL_Function_Eval

[ian-bertolacci]

In progress to #11

Current progress:

• Split the scatter section of the loop out into a new loop
• Store temporary variables into a vector
• Reverse scatter pattern into a gather pattern
• Verify output
• Check performance changes (in serial)
• Attempt parallelization

[ian-bertolacci]

I can assert that in the regular LW test case (72hr sim with outputs every 3 simulated hrs) that the difference between the baseline case (from master at commit 1de045b) compared to this branch (commit f2f7d68) in the output pressures is less than 6.5e-11 for all output time-steps, and in the output saturation is less than 1.6e-14 for all output time-steps.

[ian-bertolacci]

Whole program run times (in seconds ± 1 standard deviation) for baseline and 511_split_loop in the regular LW test case (72hr sim with outputs every 3 simulated hrs)

Time Class	baseline (master 1de045b)	511_split_loop (f2f7d68)
real	332.3808 ± 1.7961	334.5446 ± 1.3524
user	331.4917 ± 1.7961	333.6424 ± 1.3506
sys	0.7596 ± 0.0166	0.7657 ± 0.0225

Easy to see the 2.1638 second slowdown in the split-loop case.
This is most likely from the additional memory usage and the loss of caching from performing the reduction using the created data in the second loop instead of immediately when that data is created.