Solver give very different weights for n_samples < n_features

[agoscinski]

This can make sense, but it worries me a bit, so I would like to document this a bit better. Also the test test_approx_ref_numpy_solver_regularization_dual can be removed in this case.

I am still confused why the test test_approx_ref_numpy_solver_regularization_dual fails (so why the weight matrix calculated by the different solvers disagree with each other in the case n_samples < n_features), but the prediction results agree (see test test_predict), so I would think about this some other time.

Originally posted by @agoscinski in #31 (comment)