Return `gradient` with `hessian`, or `derivative` with `second_derivative`

[gdalle]

As discussed in #296, defining the following operators would allow some backends to optimize the joint computation:

• value_derivative_and_second_derivative
• value_gradient_and_hessian

Main issue: with what kind of scenario do we test those?

[adrhill]

I guess we could reuse our Hessian scenarios and add Gradient information to them?

[gdalle]

Yeah we had the same idea, after some thinking I added a first_order_ref to the second-order scenarios in #300