Learning Stable Deep Dynamics Models for Partially Observed or Delayed Dynamical Systems