upcxx::barrier() should use gex_Coll_BarrierNB()

[Dan Bonachea]

upcxx::barrier() in the 2018.9.0 release is implemented using gex_Coll_ReduceToAllNB(), which is correct but sub-optimal, and incurs a measurable performance penalty. This was done due to a semantic restriction in the 2018.9.0 GEX release, but that restriction has now been lifted in collaborator-snapshot.

upcxx::barrier() should be updated to use gex_Coll_BarrierNB() instead, which already offers better performance, and will eventually allow more aggressive h/w offload.

[Dan Bonachea]

Proposed implementation:

https://bitbucket.org/berkeleylab/upcxx/pull-requests/56

[Dan Bonachea]

Convert upcxx::barrier(_async) to use gex_Coll_BarrierNB

Resolves issue #179

Remove redundant memory fences

→ <<cset db2fe256c25e>>