Specialised state saving interface

[TorkelE]

If you wish to save things during an optimisation run (typically the loss trace, but sometimes other stuff as well) you can typically implement this during a callback and your own structure. Depending on situations this can be more or less messy.

However, it would be straightforward to have an argument to solve save_state(os::OptimizationState) -> u (potentially one would want to impose some conditions on u). Next, in each step (or every n steps), save_state is called on the optimisation state and saves the result to a vector (state_trajectory) which is saved as a field of the OptimizationSoltution.

By default, save_state would do nothing and state_trajectory would be nothing.