This fixes two buts that the glibc NPTL verification tests found, one
new and one old.

The new bug is that "offset" has been declared as an alternative in
the union, instead of as an element in the structures comprising it,
effectively eliminating it from the key: keys match which should not.

The old bug is that if futex_requeue were called with identical
key1 and key2 (sensible? tended to happen given the first bug),
it was liable to loop for a long time holding futex_lock: guard
against that, still respecting the semantics of futex_requeue.

While here, please let's also fix the get_futex_key VM_NONLINEAR
case, which was returning the 1 from get_user_pages, taken as an
error by its callers.  And save a few bytes and improve debuggability
by uninlining the top-level futex_wake, futex_requeue, futex_wait.