atomic_ref (#2, p0019r8)

[AlexGuteniev]

Fixes #2
Seems like it that simple to reuse existing atomic bases.

[AlexGuteniev]

Still requires to handle generic atomic by a lock via contention table instead of internal lock.

[AlexGuteniev]

So two main issues with this change is with mutex for bad size:

• It is hard to eliminate data member without ABI-unsafe layout change and without [[no_unique_address]] (which we probably will not have soon). So embedded _Spinlock will probably stay for atomic_ref, but will not be used.
• The mutexes for atomic_ref will have to reside in contention table, so apparently must go to a satellite DLL.

Due to the first issue it may make sense to make the best of the situation, and cache mutex reference in _Spinlock for atomic_ref. This will make (satellite) dll call only on construction, rather than on every operation.

Note that I'm assuming that embedded by value spinlock is not an option because independently constructed atomic_ref from the same object should be safe. Standard is not very clear on this though.

[StephanTLavavej]

Looks good, very minor nitpicks.

[BillyONeal]

Thanks for contributing my favorite C++20 feature :D