The Newton solver should also be able to get a Jacobian during construction

[eike-fokken]

At the moment we re-construct the Jacobian on every optimization step.
But the sparsity structure of the jacobian never changes, so we could hand in the jacobian from the previous step.
This would remove the need to re-allocate the jacobian on every optimization step.
(In code this means, we could dispense with "set_from_triplets" in all but the first optimization step).