Add a more powerful pre-processing before parsing doctypes

[lagru]

After talking with @glemaitre on what scikit-learn might need to make use of docstub, we had the idea to allow users to to regex substitution on doctests.

This would be useful to address the following doctype that scikit-learn is using in many places:

{array-like, sparse matrix} of shape (n_samples, n_features)

This should be translated to

array-like of shape (n_samples, n_features) | sparse-matrix of shape (n_samples, n_features)

which could be done with regex substitution and capture groups.

Another way would be to define a proper union type that combines array-like | sparse-matrix but we weren't sure how that should be named to be still expressive.