float_pow():  Don't let the platform pow() raise -1.0 to an integer power
anymore; at least glibc gets it wrong in some cases.  Note that
math.pow() will continue to deliver wrong (but platform-native) results
in such cases.