This version introduces small API fixes to sparse sampling.
First, we do not issue warnings anymore when constructing the
basis for "near-singular" sampling points, only if you try to
fit with it. Secondly, we now have a (..., points=...) argument
to evaluate and fit, allowing quick evaluation on different points.