Alchemy: A Benchmark Task Distribution for Meta-Reinforcement Learning Research