Optimization for move_assign and assign

[Antony Polukhin]

Current implementation of assignment and move assignment creates a temp variable if direct_assigner or direct_mover fail. We can avoid it, by checking if the type of variable to assign exactly matches one of the variant types. If it matches, we can do assign without creation of temporary variable.

Here is some pseudo-code to show it:

    template <typename T>
    void move_assign(T&& rhs)
    {
        // If direct T-to-T move assignment is not possible...
        detail::variant::direct_mover<T> direct_move(rhs);
        if (this->apply_visitor(direct_move) == false)
        {
            if (is_in_varaint_types<T>::value)
            {
                // ... but T is in varaint types, then destroy 
                // content of storage and assign
                move_assigner visitor(*this, type_to_int<T>::value);
                visitor.internal_visit(rhs, 0);
            } 
            else 
            {
                // ...then convert rhs to variant and assign:
                //
                // While potentially inefficient, the following construction of a
                // variant allows T as any type convertible to one of the bounded
                // types without excessive code redundancy.
                //
                variant temp( detail::variant::move(rhs) );
                variant_assign( detail::variant::move(temp) );
            }
        }
    }

[Antony Polukhin]

(In [86134]) Removed duplicate code in assign<->move_assign, added const modifiers (refs #7960)

[Antony Polukhin]

(In [86158]) More noexcept modifiers, especially for some of boost::get(T*) and assign/move_assign helpers (refs #7960)

[Antony Polukhin]

(In [86650]) Merge from trunk:

• dropped support of antique compilers
• fixed issue with ambiguity in swap (fixes #2839)
• added conditional noexcepts to move assignments, default and move constructors (fixes #7911)
• experimental variadic templates support (refs #9163)
• bunch of size optimizations for assignments and move assignments (refs #7960)
• minor changes and size optimizations

[Antony Polukhin]

Closing as the performance is OK now.