BRExIt: On Opponent Modelling in Expert Iteration