Merge remote-tracking branch 'origin/azamat/homme/optimize-limiter' PR …

…#992

Performance optimizations:

Linearize two nested loops into one larger np*np loop
Fuse intrinsic reduction function sum into an existing explicit loop
Vectorize the explicit linearized loop
Replace multiple array references with local variables
Fuse linearized and final update loops
Result:

before: 7% of total time in standalone NE=8 HOMME runs
after: this function is now 3.5% of total time
total run time is reduced from 7.722 seconds/sim-day downto 7.244s or 7% speedup
[non-BFB] - due to change in the order of floating-point operations