Axiomatic Preference Modeling for Longform Question Answering