Compatibility with `ad`

I have been trying to use this package with `nonlinear-optimization` and `ad` for maximum likelihood computations, however this is highly non-practical since distributions always return `Double -> Double` and `ad` needs some kind of polymorphic interface to work. Are there any barriers for such a polymorphic interface?

One problem I see is the need for different contraints depending on the distribution (`Erf` for Normal, `Num` for others).