Cleaner stochastic solvers

[stephanegaiffas]

Adding some extra hierarchy in C++ stochastic solvers and some extra tests in computations. For instance, some methods do not check that a model was actually passed, and that the model is indeed a GLM... This could lead to segfault (if really done on purpose...). Also, some things could be mutualized, such as some precomputations of step-corrections (for solvers accepting sparse datasets).

[Mbompr]

Is this a bit like #94 ?