ENH improve integration with encoder API

[tomMoral]

Things that we can do/discuss for alphacsc/dicodile integration. Let's discuss those on monday.

• Add tests for the z_encoder behavior and the different backends.
• Implement a first dicodile powered Encoder backend
• Overall, avoid calling get_z_hat if not strictly necessary. This is costly in communication and all operations that can be distributed are more efficient.
• In compute_d_multi, give z_encoder instead of X, z_hat, D_hat, constants. This way, depending on the algorithm, it can either call get_sufficient_statistic or get_z_hat depending on what it needs. Maybe also we can see if we can implement some other operations in update_d_multi this way.
• For callback, also try to use the z_encoder as much as possible.
• implement efficient operation when D_hat is composed of rank_1 atoms in dicodile
• implement add_one_atom in dicodile
• implement get_error_max in dicodile in a distributed algorithm. (see Implement get_max_error_patch in dicodile tomMoral/dicodile#49 ).
• implement a multi signal dicodile version for online learning.

[rprimet]

One missing step: Implement a first dicodile-powered Encoder backend?

At which point should we do that? (maybe should be done early on, before optimizations, in order to troubleshoot/complete the Encoder API?)

[tomMoral]

Yes Indeed, I updated the list of tasks.