On the Geometry of Learned Representations in Event-Based Multi-Modal Egomotion Estimation