Decoupling Meta-Reinforcement Learning with Gaussian Task Contexts and Skills