Before this patch raising a big-int to a negative power would just return
0.  Now it returns a floating-point number with the correct value.