Gotta Learn Fast: A new benchmark for generalization in RL