Fix potential false sharing in ReduceOps.eval with OMP

[AlexanderSinn]

Summary

Looking at the code I noticed that eval() with OMP could be prone to false sharing.
TBD if this is actually the case and if this PR improves it

Additional background

Checklist

The proposed changes:

• fix a bug or incorrect behavior in AMReX
• add new capabilities to AMReX
• changes answers in the test suite to more than roundoff level
• are likely to significantly affect the results of downstream AMReX users
• include documentation in the code and/or rst files, if appropriate

[AlexanderSinn]

Initial benchmarks show only a very small difference, if any

[WeiqunZhang]

In any way, it will not hurt.

[zingale]

this changes answers in Castro's CI -- is that to be expected?

[AlexanderSinn]

A change in the order of machine precision is expected due to floating-point addition not being associative to that precision.