Reasoning in visual navigation of end-to-end trained agents: a dynamical systems approach