Consider explicit conversions before aggregate fallback

[yronglin]

Test:

struct float2 {
    float x;
    float y;
};

struct __nv_fp8x2_e4m3 {
    float x;
    float y;

    explicit operator float2() const {
        return float2{x, y};
    }
};

float2 foo(__nv_fp8x2_e4m3 val) {
    float2 a(val);
    return float2(val);
}

clang and gcc accept this code but edg reject it:

"<source>", line 16: error: no suitable conversion function from "__nv_fp8x2_e4m3" to "float" exists
      float2 a(val);
               ^

"<source>", line 17: error: no suitable conversion function from "__nv_fp8x2_e4m3" to "float" exists
      return float2(val);
                    ^

[over.match.ref]/1.1: when binding the const float2& parameter of the copy constructor to the argument, "For direct-initialization, the permissible types for explicit conversion functions are...", the explicit operator float2() const is therefore a legitimate candidate:

• float2(const float2&) is viable.
• the precondition "if no constructor is viable..." in [dcl.init.general]/15.6.2.2 does not hold, so the aggregate fallback branch must not be entered.

https://godbolt.org/z/oKoMvEEdd