Ramp coding with population averaging predicts human cortical face-space representations and perception