Should all models match the sklearn interface?

[jcrist]

Currently most models in cuml match the sklearn interface (or at least are intended to). The exception seems to be the models in cuml.tsa, which match the interface in statsmodels IIUC.

• Are we ok with different interfaces being implemented in this project (my guess is yes)?
• If so, should we use a separate base class than cuml.internals.Base for those that don't match the sklearn interface?