fix(libcudacxx): Temporarily bypassing managed memory for atomic op
临时绕过原子操作的统一内存用法，暂未支持。以下三个例子相关。
heterogeneous/atomic.pass.cpp
heterogeneous/atomic_ref.pass.cpp
heterogeneous/cuda_atomic_ref.pass.cpp
http://hpczentao.sugon.com/bug-view-114264.html