mutex(9): Convert to membar_acquire/release.

Except for setting the waiters bit -- not sure if this is actually
required to be store-before-load/store.  Seems unlikely -- surely
we'd have seen some serious bug by now if not, because membar_enter
has failed to guarantee that on x86! -- but I'm leaving it for now
until I have time to think enough about it to be sure one way or
another.

(riastradh)