teach GCC how to use CMPXCHG on 128 bits

On x86-64, the reference counted pointer implementation ends up being a 128-bit
scalar (`__int128` provided by GCC, Clang and ICC). The MOV instruction on
x86-64 is only guaranteed atomic up to 64 bits, a quadword. To get atomic
operations on 128 bits we need to leverage CMPXCHG. The result is a counter-
intuitive degenerate CMPXCHG to implement an atomic read.

A further complication is that GCC version >= 7.1 will not emit CMPXCHG for
__atomic_compare_exchange* even when you pass the command-line flag '-mcx16'.
To get around this, we use the old __sync built-ins, despite that they are
considered deprecated. For more information, see
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80878.

We don't consider this to fully close Github #32, because a further optimisation
would be to consider simply doing the initial read non-atomically. In a sense,
it doesn't matter if we perform a torn read here because we only commit it in a
CMPXCHG, which would simply fail if the dependent read was torn.

Related to Github #32 "mcx16 hacks"