As well as `RUBY_ATOMIC_PTR_EXCHANGE` and `RUBY_ATOMIC_PTR_LOAD`.

Use `_WIN64` for word-size, `_M_AMD64` for CPU-specific feature.

This can already be emulated by doing an atomic fetch_add of zero, but
this is more explicit.

[Bug #19994]

On MinGW, where both of Win32 API and GCC built-ins are available,
the mismatch of implementations to the definition caused lots of
warnings.

Reduce macros to do the same things in inline functions instead.  This
way assertions can be made granular.

Must not be a bad idea to improve documents. [ci skip]

Pointed out by @shyouhei.

NOTE: Already we have dropped the support for older MSVCs,
probably prior to 1300 or 1400.  Remove the conditional code,
especially in win32/Makefile.sub.

Now we need atomic operations, which are lighter than mutex, more
widely for extension libraries because of Ractor.