pow broken for 0^(-x)

[Pflugshaupt]

#include <iostream>
#include "xsimd/xsimd.hpp"

int main(int argc, char *argv[]) {
  std::cout << xsimd::pow(xsimd::batch<float>(0.f), xsimd::batch<float>(-1.f)) << "\n";
  std::cout << std::pow(0.f, -1.f) << "\n";
}

produces (on sse2+ and neon)

(0, 0, 0, 0)
inf

The reason is the fix applied in #684. The result is always set to zero for input zero to work around floating point exceptions, but it should be infinity for 0^(negative number). I worked on a patch, but wasn't clever enough to do this without adding an extra comparison and an extra select(). This feels like it should be possible with one clever check somehow.

[serge-sans-paille]

\o/ managed to do it with an extra comparison + a logical and, no extra select.