don’t implement cmplx with +

For brevity, let’s forget that CMPLX is a macro. The following code seems to
be a cross-platform solution.

inline double _Complex CMPLX(double x, double y)
{
	return x + I * y;
}

Actually, it is not. Whether I is complex or imaginary, I * y evaluates to
(+0, y) when added with a real number. If x happens to
be -0, its sign is not preserved because -0 + +0 = +0.

I think we can only stick with compiler-specific constructs for now. 🙁