End-to-End Reinforcement Learning of Koopman Models for Economic Nonlinear Model Predictive Control