Enhance the performance of fn_ref.

[Kim-J-Smith]

In previous implementation, both fn_ref and fn will pass the address of m_erasure when calling operator(). However, it is actually an ODR use of m_erasure, which means that the compiler CANNOT optimize the object stack space of fn_ref.

In new implementation, erasure_type::ErasurePass is added as the new first parameter for the ABI of m_invoker to take place of erasure_type::ErasureBase*. This allows operator() to pass m_erasure by value in fn_ref to avoid ODR use of the object.

Here is the assembly difference.

• old(GCC 16.1 -std=c++26 -Os) https://godbolt.org/z/EadE5813z :

"check1(ebd::detail::function<8ul, ebd::detail::fn_traits::config_package<true, true, false, false>, bool (int, int)>)":
        sub     rsp, 24
        mov     edx, 3
        mov     QWORD PTR [rsp], rdi ; extra stack frame and indirection
        mov     rdi, rsp
        mov     QWORD PTR [rsp+8], rsi ; extra stack frame and indirection
        mov     esi, 2
        call    [QWORD PTR [rsp+8]]
        add     rsp, 24
        ret
"check2(std::function_ref<bool (int, int)>)":
        mov     rax, rdi
        mov     edx, 3
        mov     rdi, rsi
        mov     esi, 2
        jmp     rax

• new(GCC 16.1 -std=c++26 -Os) https://godbolt.org/z/Yrzsvz3G9 :

"check1(ebd::detail::function<8ul, ebd::detail::fn_traits::config_package<true, true, false, false>, bool (int, int)>)":
        mov     rax, rsi
        mov     edx, 3
        mov     esi, 2
        jmp     rax
"check2(std::function_ref<bool (int, int)>)":
        mov     rax, rdi
        mov     edx, 3
        mov     rdi, rsi
        mov     esi, 2
        jmp     rax