Add generic support for Sigmoid.

This maintains the existing logic of doing the math in
double precision and converting back to the intended
type (previously: just float).  We do the same for
half here, although perhaps we should do the math
at float in that case.

There is some question about what to do with conversions;
Sigmoid did math in double before converting back to float;
we keep this intent, although there is some question on whether
this was intentional and for half -- should we just go up to
float or up to double?