Fix CUDA architecture dependent issues
Only device code gets generated in multiple passes and therefore
target architecture-dependent macros like __CUDA_ARCH__ or our own
IATYPE_SHMEM (which also depends on __CUDA_ARCH__) are not usable in
host code as these will be both undefined. As a result, current code
over-allocated dynamic shared memory. This has no negative side-effect.
This change replaces the use of macros with runtime device compute
capability checks. Also texture objects are now actually enabled,
which give very minor performance improvements.
Note that on Maxwell + CUDA 7.0 there is a 20% performance regression
for the tabulated Ewald kernel (which is not used by default), which
magically disappears when texture references are used instead.
Change-Id: I1f911caad85eb38d6a8e95f3b3923561dbfccd0e