use __sync builtins for dword atomics on all architectures

The Debian buildd logs for the last arm64 build [0] indicate that
NEEDS_LIBATOMIC was False, but later linking then failed with a missing
reference to __atomic_load16 and friends. I'm guessing this is something like
the x86 scenario, where only the __sync built-ins are capable of producing
inline code. Let's unconditionally use them on all platforms (not just x86 and
x86-64) and see if that resolves these issues.

  [0]: https://buildd.debian.org/status/fetch.php?pkg=rumur&arch=arm64&ver=2020.04.26-1&stamp=1588101624&raw=0

Github: related to #184 "debian: arm64 broken build"
Github: related to #185 "debian: armel broken build"