Auto-diff

This version comes with support for Sparse Jacobians using `SpareAutoDiff`.

`SparseAutoDiff` can be used exactly like `AutoDiff`, except users must not use the `np.ndarray` constructor in their code or in libraries they call into. We recommend `np.empty` instead.

Fixes bugs in the inverse hyperbolic trigonometric functions and with `VecValDer.__set_item__`.

Adds all of the Hyperbolic Trigonometric Functions.

This release fixes a number of bugs with broadcasting in our ufunc implementations for

* `np.add`
* `np.subtract`
* `np.multiply`
* `np.divide`
* `np.power`
* `np.float_power`

* Fixes a bug with `np.float_power` and `np.power`
* Adds more methods to `VecValDer`