Bug fix and simplification for CUDA X Buffer Ops

Bug fix and simplification for CUDA X Buffer Ops

Fixes bug when there are more than two grids (ie they don't naively
map to local/nonlocal), by using a separate GPU memory space for
cxy_na and cxy_ind data for each grid (it previously only had one
local and one nonlocal space with overwriting). Also simpifies by now
only calling init fn once per NS step to set up data for all grids
(previously called for both local and nonlocal).

Change-Id: Ia2b97d22324aa97dca34b05da2eca2e2090372af