197 pytree regularizer strength

[wulfdewolf]

This extends regularization by allowing passing pytrees that match the parameter structure.
Every parameter can thus be regularized differently.
This will be particularly useful for the PopulationGLM, as it allows regularizing individual neurons differently.

Some details:

• For the moment intercepts are not regularized (as before).
• ElasticNet regularization now also takes a pytree (strengths, ratios are set to 0.5) or a tuple of pytrees (strengths and ratios), to allow for maximum flexibility.

TODO:

• ElasticNet penalization
• ElasticNet proximal operator
• GroupLasso penalization
• GroupLasso proximal operator
• tests
• docs