Text this: VariBAD: a very good method for Bayes-adaptive deep RL via meta-learning