Complex pow numerics at zero

[axch]

pow (MkComplex 0. 0.) (MkComplex 2. 0.)

produces NaNs, even though the answer is 0.

Should we

• Special-case zero base to return 0? In this case, we should also special-case 0^0 for complex numbers to be 1? Such special cases are relatively cheap on CPU, because CPUs swing through branchy code like monkeys; but a GPU interacts with branches more like an elephant.
• Special-case integer powers to use multiplication (maybe with some type-level trick to avoid runtime dispatch)? (But can we even implement the O(log(power)) algorithm?)
• Leave it the way it is and let users work around it themselves?