MALT: Improving Reasoning with Multi-Agent LLM Training