Effective Distillation of Table-based Reasoning Ability from LLMs