Add (Prox-)SVRG as a solver for GLMs

[bagibence]

Add a preliminary implementation of SVRG and Prox-SVRG compatible with the jaxopt API, so nemos can use it as a solver:

• the run method (used in GLM.fit) assumes the full data is loaded into memory and runs optimization until convergence.
• the update method (used in GLM.update) updates the parameters on a single data point or batch. Note that currently, when fitting on batches and calling glm.update, some steps of the algorithm have to be implemented outside.

ProxSVRG takes a prox argument for the proximal operator which should be a callable with the same signature as the functions defined in jaxopt.prox (e.g. jaxopt.prox.prox_lasso).
SVRG inherits from ProxSVRG, uses the identity as the proximal operator and adjusts some calls to account for the different number of arguments.

[billbrod]

This generally looks very good! Mainly some minor changes here, and I responded in a couple of my previous comments, so make sure to check those again.

Also, and I know this is annoying, but test_solvers still has references to xs and df_xs. I think that should be updated to match the new nomenclature, which will make it easier for us in the future.

[billbrod]

Two small issues and then this is ready to merge!

• Docstring of ProxSVRG mentions hyperparams_prox, which is no longer an argument. So that should be removed.

[BalzaniEdoardo]

• hyperparams_prox

The docstrings are ok, because that is referring to any hyperparameter that a general proximal operator gets as input. For example, what we call regularization_strength is one such hyper-parameter for ridge.