End-to-End Stable Imitation Learning via Autonomous Neural Dynamic Policies