Investigate potential use of `StochasticAD.jl` for non-differentiable models

[pat-alt]

It might be possible to use StochasticAD.jl (see paper) to differentiate models that do not have a continuous dependence on features (such as decision trees).

[pat-alt]

@mschauer haven't carefully studied your paper yet, so not sure if I'm being overly optimistic here - do think this could work?

[mschauer]

Probably possible to find models, where it works

cc @gaurav-arya