Currently the ridge with generalized cross-validation which uses an SVD of the
design matrix (the best when there are more samples than features) does not
support sparse design matrices.

this results in silently using an eigendecomposition of the Gram matrix when
gcv_mode is 'auto'

or raising an error when the user explicitly asked for svd

would it be better to:

• allow sparse design matrices and compute the SVD using scipy.sparse.linalg.LinearOperator

• warn the user that 'eigen' is being used, which can be very inefficient when
  n_samples is much larger than n_features, so that they can decide if it is
  better to use 'eigen', cast X to a dense matrix, or turn to another mode of CV
  than the generalized CV