ENH add optinal torch-based solver

[tomMoral]

In our recent paper with @Malezieux, we achieved faster performances with almost no performance drop with a torch-based algorithm.
It would be nice to implement such algorithm in alphcsc, with an optional dependency on pytorch.

Where to add this algorithm is not totally straightforward. I think the easier would probably be to implement a separate algorithm similar to _batch_learn and _online_learn. But another possibility would be to refactor the code to better take into account the bi-level optimization framework (less alternate minimization), which would ease this integration (and follow up that will unfold from our work on the field).

[agramfort]

we need to organize a code sprint ....

…

Message ID: ***@***.***>