Learning Topology-Agnostic EEG Representations with Geometry-Aware Modeling