Add function to convert internal derivatives to external ones

[janosg]

Terminology

Assume we have a criterion function f(params) and its derivative (gradient or jacobian) d.

To enforce constraints, the internal optimizer sees a reparametrized version g(x) = f(r(params)).

By internal derivative we mean the derivative of g with respect to x.

Current situation

We have closed form derivatives of the reparametrization function r and can convert external derivatives to internal ones.

Desired Situation

We also want to convert internal derivatives to external ones, for example when we only calculate an interal derivative numerically but would want to store an external one in the log file.

Proposed implementation

• implement closed form derivative of unscaling (trivial)
• calculate numerical derivatives of reparametrize_from_internal
• combine everything, similar to _derivative_to_internal in conversion.py.

Alternative implementation

Alternatively, we could implement closed form derivatives fo reparametrize_from_internal but I do not think it's worth the work.