OWL: Geometry-Aware Spatial Reasoning for Audio Large Language Models