Adding Autodiff for predictions

## 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