Change the `nbconvert` test so that it converts a notebook with no Hoogle (which changes over time) and no images (which differ on different systems). This should reduce the number of surprise Travis test failures.
