`voigt` could be more accurate

At present, we are using the approximation from Hunger 1965, which is the same at MOOG. Previously, I thought there was a typo in the source, but it's looking more likely that the inaccuracies are just a limit of the approximation. Unfortunately, this means that our voigt function breaks my 1% precision "rule". At some point, we should look into doing better.

Note that in "normal" part of parameter space, where the Lorentz width is less than the gaussian width, there is no problem. So, maybe this doesn't need to be fixed. To put this to rest, I would at least like to verify that we aren't hitting that part of parameter space for any realistic problems.

(The parts of parameter space not shown are fine.)

Another perspective on the same thing. This shows the computed profile for alpha in `1 : 0.2 : 2`. (Korg voigt in red, true voigt in black).