Large Language Model Cascades with Mixture of Thoughts Representations for Cost-efficient Reasoning