In the sparse case, Ridge silently ignores fit_intercept (currently, it's equivalent to fit_interept=False).

Using the recently added add_dummy_feature, it should now be easy to support fit_intercept=True.

Since it corresponds to penalizing the intercept, we need to add a intercept_scale option to the constructor (like LinearSVC).

add_dummy_feature results in a copy of X. I suggests to set fit_intercept="auto" by default. In the dense case, auto can correspond to fit_intercept=True and in the sparse case to fit_intercept=False (to avoid the memory copy). This way, the code will be backward compatible too.