Support for composition of nonlinear functions with affine arguments

[Transurgeon]

@dance858 and I are working on a differentiation engine which will support taking derivatives of compositions of nonlinear functions with affine arguments of the form $\phi(Ax)$.

Currently, we don't support any compositions (this is on the canonicalized problem), and can only perform derivatives of $\phi(x)$ where $x$ can be a newly introduced variable (with the equality constraint $x=args$).
This can potentially become very costly, especially as the number of expressions grows (i.e. some unvectorized formulations).

It would be good to have some sort of working prototype on the DNLP side before integrating in the differentiation engine.
One way I see to do this is to sort of collapse the affine subtrees (that are arguments to nonlinear functions) into one $A$ matrix. This is what will be required to initiate the differentiation engine.