Knowledge-driven Scene Priors for Semantic Audio-Visual Embodied Navigation