Make stochastic solvers fit in this benchmark

Now that we use `SufficientProcessCriterion` to stop the benchmark and that we can have multiple loss at once, there is no real gain to have separate benchmarks for `logreg_l2` with batch and stochastic solvers. We should thus merge benchopt/benchmark_stochastic_logreg_l2 and this benchmark. To do this, we should:

- [ ] port `SGD` solver in this benchmark
- [ ] add a notion of `train/test` losses in `objective.py`
- [ ] improve the plotting utils for multi value loss functions to get train/test graphs.
- [ ] close the benchopt/benchmark_stochastic_logreg_l2 repo

Moreover, to make this benchmark more inline with literature practices, we should try and reproduce the following figures:

* From [Schmidtt et al. 2014](https://arxiv.org/pdf/1309.2388.pdf)

* From [Leblond et al. 2017](http://proceedings.mlr.press/v54/leblond17a/leblond17a.pdf)

Compare benchopt/benchmark_stochastic_logreg_l2 with this benchmark first, then inspect objective.py and the plotting utils. Port the SGD solver, represent train/test losses, produce train/test graphs, reproduce the referenced figures, and close the separate stochastic benchmark repository when the merged benchmark is complete.

Written by the indexing model from the issue text.