Improve overload resolution to support P0608

[jbuonagurio]

This PR is an initial attempt at adding support for P0608R3: "A sane variant converting constructor" (https://wg21.link/p0608) so conversions like this should work as expected:

namespace variant2 = boost::variant2;
variant2::variant<bool, std::string, int, int> v("test");
assert(variant2::holds_alternative<std::string>(v));

The tradeoff is that user-defined types convertible to bool may construct a different type. More information in P1957 (https://wg21.link/p1957)

I'm fairly new to metaprogramming so any feedback on the implementation and/or correct use of Mp11 is much appreciated. I haven't yet added workarounds for the older MSVC versions and it's not yet thoroughly tested, but I can help write some test cases if you're interested in merging.

[pdimov]

The reason your is_arithmetic check doesn't seem to work is that mp_invoke_q is immediately evaluated regardless of the value of is_arithmetic. It should probably be

template<class T, class U> using narrowing_check = mp11::mp_eval_if<std::is_arithmetic<T>, mp11::mp_identity<T>, mp11::mp_invoke_q, narrowing_check_impl, T, U>;

However, this check doesn't seem necessary; the tests are just wrong for a C++20 variant. E.g. https://godbolt.org/z/KMcnb1