Adding Autodiff for predictions

[odunbar]

Issue

In some cases, once trained the predicted mean and covariance will be differentiable. Having a differentiable regressor can be useful in situations where the regressor is used for optimization or sampling etc. How can we enable this functionality

Possible solutions

As the RFs are simple objects, akin to neural networks, the ForwardDiff.jl/Zygote.jl packages for forward/backward differentiation may be used to differentiate the output, and are expected to work almost out-of-the box. We could investigate this