internal/refcount.h: overhaul fencing and add _MSC_VER section.

Relax memory_order on counter decrement itself, because mutable
members of the reference-counted structure should be visible on all
processors independently on counter. [Even re-format and minimize
dependency on other headers.]

Reviewed-by: Kurt Roeckx <kurt@roeckx.be>
(Merged from https://github.com/openssl/openssl/pull/6900)