Optimize any::swap

[achabense]

This pr removes an unnecessary copy in any::swap.

Hopefully fixes #5698.

Benchmark result:

------------------------------------------------------
Benchmark            Time             CPU   Iterations
------------------------------------------------------
bm<trivial>       14.5 ns         14.8 ns     49777778
bm<small>         21.1 ns         21.0 ns     32000000
bm<large>         15.2 ns         15.0 ns     44800000

->
------------------------------------------------------
Benchmark            Time             CPU   Iterations
------------------------------------------------------
bm<trivial>       10.1 ns         9.52 ns     64000000
bm<small>         15.9 ns         15.3 ns     40727273
bm<large>         10.6 ns         10.7 ns     64000000

[achabense]

We are still paying an additional copy for some insane self-swaps here...

std::any a = std::make_any<std::any>(42);
std::any& a_o = *std::any_cast<std::any>(&a);
a.swap(a_o);
a_o.swap(a);

(Also, I think _Assign(any-prvalue) is somewhat misleading (for example #5413)), and it's better to create the temporal any explicitly in each function. For example, this would be v.)

any _Old_this = _STD move(*this);
{
    any _Old_that = _STD move(_That);
    reset();
    _Move_from(_Old_that);
}
_That.reset();
_That._Move_from(_Old_this);

[achabense]

The benchmark is very simple, but I think it's enough to show the compiler cannot optimize away the copy.