ENH fit_intercept for sparse design matrices

This is a follow-up to PR #94, where fit_intercept support was added for dense data only in some solvers.
For some of these solvers fit_intercept support can be extended to sparse matrices with reasonable effort.
This issue will be used to keep track of this enhancement.

- R-PGD: workaround to [pass a sparse matrix via rpy2](https://stackoverflow.com/questions/48488665/running-glmnet-with-rpy2-on-sparse-design-matrix)
- CD: the code could be directly modified to compensate for X_offset. So, X would not be explicitly centered in order not to break sparsity.
- (...) To be continued: list other solvers that can be adapted and how.

Start by reading PR #94 and the solver notes in this issue, especially the CD and R-PGD approaches. Identify which solvers can support fit_intercept with sparse design matrices without breaking sparsity; done means the selected solvers handle this case and have coverage for the new behavior.

Written by the indexing model from the issue text.