Cleanup up handling of references.

[joto]

Storing references hasn't worked before. This commit documents the fact, adds a
static check for it and tests this case. We also test the recommended
alternative, which is to use std::reference_wrapper.

[springmeyer]

👍 This looks good.

[lightmare]

I think the last diff hunk is wrong. As is explained in the link in the comment above it: since T is a deduced type, T&& is a universal/forwarding reference, it can bind to rvalue reference or lvalue reference. T may be deduced as X const& for some type X, that's why remove_reference is needed -- because (this implementation) variant can only store X, not X const&.

[artemp]

@joto - @lightmare is right and that was my motivation: to allow T being deduced either to rvalue (move) or lvalue (copy). I'll take a fresh look tomorrow again but the original code looks correct to me.

[lightmare]

A concrete example for illustration:

int a = 5;
variant<int> v(a);

The conversion constructor will be variant::variant(T && x) with T == int&. remove_reference<T> == int, that's what you want to store.

[joto]

As per chat with @artemp I committed the changes to master without the wrong removal of std::remove_reference and close this pull request now.