(Free42) roundoff for complex SQRT

[Thomas Okken]

If I'm going to check for Re(X)=0, cases where Im(X)/2 is a perfect square should return exact results, or at least no worse than SQRT for real arguments.

That 48G code doesn't check for pure-real or pure-imaginary arguments, but Free42 uses such checks in lots of places, for various reasons, so it doesn't bother me to add one to complex SQRT as well.