Reliable Learning and Control in Dynamic Environments: Towards Unified Theory and Learned Robotic Agility