Work around bugs in COMPLEX-ERF. Bug #1: sqrt(z^2)/z incorrect
when Re(z) = 0 and Im(z) < 0. Bug #2: (EXPT Z 2.0) yields spurious imaginary part
in some Lisps. Bug #3: GAMMA-INCOMPLETE returns different values (conjugates) for
imaginary part of z^2 equal to 0.0 or -0.0, if the Lisp implementation recognizes
Resolution of #1 and #3 is to work in upper half plane. Resolution of #2 is to
replace (EXPT Z 2.0) with (* Z Z).
Incidentally need to increase tolerance for erf(-0.75+%i) test as (* Z Z) yields
slightly different results for some Lisps.