reverted to L2 regularization for splined models

The difference-from-mean regularization wasn't good as a "regularization". Any
place would be a perfect fit as far as the regularization is concerned. So the
optima found by the solver would end up with giant focal lengths, compensated
for by crazy intrinsics