There are multiple issues with that commit, incl. out-of-bounds read
on x86 systems and a significant performance decrease for qualified
floating-point types; see #8124 for a full list of flaws.
I'm questioning the tackled goal in the first place, because I expect
users to use that `pow` mainly with small integral exponents, *not*
expecting underflows to 0 or overflows to ±infinity, so optimizing
for these cases and sacrificing performance for the regular case doesn't
make a lot of sense to me.
