Error in preprocessing pipeline may go undetected until after search

[PGijsbers]

With the minimal example of #131 an error is thrown after search, even though the error arises in the (prepended) preprocessing pipeline. Any error generated by the preprocessing steps that are determined globally should be caught before search.
There are multiple ways to catch this early, but a fit_transform on the preprocessing pipeline is probably sufficient for now.
Ideally we cache the outcome but that should probably be added only after revising the compute back-end (#3).