Enabling meta learning in Shogun

This page contains the logs of things that have been and will be worked on to enable meta learning in Shogun (collaboration with the ATI).

First part:

Update the Shogun codebase to use a single method to register parameters that can then be observed using the Reactive model.

• Refactor AnyParameter.h to have a clean differentiation between model parameters (weights, bias,...), hyperparameters (k in KMeans, regularisation parameters,...) and gradient parameters (#4412).
• Refactor SGObject to add all parameters with the new API (#4417)