Sample Complexity of Reinforcement Learning using Linearly Combined Model Ensembles