[CUDA] Added stubs for __nvvm_atom_add_*_d() builtins.

They show up in CUDA headers but are not currently supported by
NVPTX back-end.

Differential Revision: http://reviews.llvm.org/D11694

llvm-svn: 245502