Update QN operator model

[tmigot]

Connected to #179
Here we update the QN operator after solving the penalized problem.

Can we do more?

[tmigot]

@MaxenceGollier I am tagging you here because I think you were interested in this.

[github-actions]

Package name	latest	stable
JSOSuite

[dpo]

In nonsmooth solver R2N, @MaxenceGollier implemented the quasi-Newton update as a callback. I think that’s a good idea and that all solvers should probably do the same. We would provide a default callback that performs the standard update. One reason why it would be useful in that some applications have an objective of the form $f + g$ where we have $\nabla^2 f$ but $\nabla^2 g$ must be approximated.

Maybe we could start doing that here in Percival?

[MaxenceGollier]

I can do that.
Perhaps it would make sense to implement this callback in SolverCore.jl and then let the solvers of the organization use it ?

We could also add a unified linear algebra interface there instead of having to reimplement one for each solver, just a suggestion.

Let me know what you think.

[dpo]

The default callback will basically just call push!() with the operator, and all that is already defined in LinearOperators. Let’s please implement it here and see what it looks like before making a decision.