remaining issue on the mljinterface

[pasq-cat]

          @Rockdeldiablo I've added the following changes now to make things work without having to overload the `update` method or changing it in MLJFlux:

• The train method now returns la, optimiser_state, history where la is the Laplace object. This way, the object does not need to be stored as a field of the struct and the problem with update is avoided.
• To facilitate this change, calling a Laplace object on an array, (la::AbstractLaplace)(X::AbstractArray) now simply calls the underlying neural network on data. In other words, it returns the generic predictions, not LA predictions.
• The fitresult method was adjusted also for the classification case.

Now that tests are passing, there are a few more things to do (possibly in a new issue + PR) if you like.

• Add a (short) tutorial to the documentation.
• Double-check if the code in src/mlj_flux.jl can be streamlined further (e.g. do we actually still need to overload MLJFlux.build)?

For now, feel free to to focus on the other PR, just ping me and @MojiFarmanbar when you come back to this one. I need to move on to other things for now.

Originally posted by @pat-alt in #92 (comment)

[pasq-cat]

i have been trying to cross check the various branches and i think that there is a change now that has to be implemented.

• the new parameter ret_distr has to be added to the struct.
• predict has to be changed both for classification and for regression

[pasq-cat]

I will work on the docs in a pull request separate from the coding part if you don't mind.