Inconsistency between univariate and multivariate models.

[TomDLT]

⚠️ The API is not consistent between the different models. ⚠️

Models:

• (1) Univariate CSC (learn_d_z)
• (2) Univariate CSC with alpha-stable noise (learn_d_z_weighted)
• (3) Multivariate CSC (learn_d_z_multi or BatchCDL) (with or without rank-1 constraint)

API differences:

• (1) and (2) use z_hat.shape = (n_atoms, n_trials, n_times) while (3) uses z_hat.shape = (n_trials, n_atoms, n_times)
• (1) and (2) do not have a class API.
• (1) and (3) use different parameter name (ds_init and D_init)

Also, (1) and (2) do not implement the following features

• refitting z_hat on a frozen support for unbiasing
• temporal window re-parametrization
• chunk initialization strategy fc38373
• scaled regularization parameter dfe1b7b
• LGCD solver for the Z-step
• greedy algorithm, online algorithm
• restart of zero-activated atoms
• atom sorting by variance

A solution would be to make the functions (1) and (2) private, to ease the use of (3) in the univariate case, and to make a new class for model (2) (using internally (3) instead of (1)).