Near-Minimax Optimal Classification with Dyadic Classification Trees