Improving Tree-Structured Decoder Training for Code Generation via Mutual Learning