Design Matrix as Jacobian

[cortner]

Something we should seriously look into: If a model is linear, then the design matrix is simply the hessian of the loss or jacobian of the residual vector. If we manage to compute this with similar efficiency as what we do in ACEpotentials now, then this would be an extremely convenient and elegant way to work with linear models. (and it would generalize to nonlinear models as well...) Most likely this would require implementing frules for everything.