Influenciae

Fix the implementation of `RPS_LJE` to follow closely the method as described in the paper. This includes a new abstraction for the techniques derived from the Representer Point Theorem.
Add and improve the documentation on some of the newer methods that were introduced in previous releases.
General bug-fixes.

Introduce boundary-based influence values: attach a notion of how influential a data-point depending on its distance to the decision boundary, or how much the weights must be perturbated to make the model predict differently.

Add an implementation of the LiSSA algorithm for efficiently computing IHVPs without instantiating the hessian matrix.

Implement the Arnoldi algorithm for efficiently computing influence values on very large models.

Fix some problems with top-k computations with the RPS L2 implementation and add a stabilization constant for the divisions.