Decision-Theoretic Meta-Learning: Versatile and Efficient Amortization of Few-Shot Learning