Revert "use __sync builtins for dword atomics on all architectures"

This reverts commit b9c98f1a. It seems the
__sync CAS built-ins are not implemented on dwords for arm64 and s390. So we end
up setting NEEDS_LIBATOMIC=True, but then still failing compilations. Here we
switch back to the old behaviour and also modify the NEEDS_LIBATOMIC test.