On the Practical Consistency of Meta-Reinforcement Learning Algorithms