[CUDA] Implemented __nvvm_atom_*_gen_* builtins.

Integer variants are implemented as atomicrmw or cmpxchg instructions.
Atomic add for floating point (__nvvm_atom_add_gen_f()) is implemented
as a call to an overloaded @llvm.nvvm.atomic.load.add.f32.* LVVM
intrinsic.

Differential Revision: http://reviews.llvm.org/D10666

llvm-svn: 240669