LARK: Learnability-Grounded Trajectory Selection for Efficient Reasoning Distillation | ArxivCSExplorer