Physics-Driven Spatiotemporal Modeling for AI-Generated Video Detection