Group-lasso

-------------
* Improved scikit-learn compatibility
* Fixed bug where fitted estimators were unpicklable

-------------
* Fixed so GroupLasso objects can be initialised without specifying the groups
* Fixed bug with viewing the selected groups
* Changed to ConvergenceWarning instead of RuntimeWarning for when FISTA doesn't converge
* Fixed typo in warning message
* Fixed bug with list input for the groups
* Added matplotlib to requirements.txt
* Changed how groups are parsed, should hopefully fix some unreported bugs

-------------

* Fixed bug with how LogisticGroupLasso applied intercept in some methods
* Implemented backtracking line search for Lipschitz
* Added centering-based preconditioning for dense design matrices

-------------

* Fixed bug where the old_regularisation and supress_warning arguments could only
be set by the base group lasso class, but none of the actual models.

-------------

* Fixed bug where the regularisation was multiplied with the Lipschitz coefficient
of the loss gradient.

-------------

* Merged logistic group lasso and multinomial group lasso to one class.
* Added support for sparse matrices.
* Added utility for extracting groups from a sklearn OneHotEncoder.
* Added option to control group-wise scaling of regularisation.