Roll our own Jacobians

[fjebaker]

We can provide Jacobian functions for the optimizers. The benefit in doing this is we know ahead of time some elements of sparsity in the Jacobian, and can use the reflection to construct a task-graph for calculating it. We can also then do things like blend AD and finite diff.