Improve Atomics Performance on KNL/Broadwell where PREFETCHW/RFO is Available

[nmhamster]

Initial research studies show there may be improved atomics performance on machines with RFO/PREFETCHW instructions available assuming atomics have low contention rates. Want to investigate adding this to Kokkos atomic operations to improve performance.

[crtrott]

Did someone do any performance analysis (i.e. when does it help when does it hurt?) In the end most of the use cases are probably low contention so I guess its a good change overall with little downside. But we should document this somewhere.