Complete r323772 on amd64.

Compilers are allowed to combine plain reads into group operations,
e.g. 64bit element copies of one array into another can be
legitimately optimized back to a memcpy() call, which r323772 tried to
prevent.

Qualify accesses to LDT descriptors with volatile dereference to
ensure that each write indeed occurs. After that, our usual claim of
native-size aligned writes being atomic applies.

This is equivalent to atomic_store(memory_order_relaxed) C11 accesses,
but our machine/atomic.h does not provide corresponding primitive.

Noted and reviewed by: bde
Sponsored by: The FreeBSD Foundation
MFC after: 1 week