Grounding Spatial Relations in Text-Only Language Models