From Large to Tiny: Distilling and Refining Mathematical Expertise for Math Word Problems with Weakly Supervision