Can Small Language Models be Good Reasoners for Sequential Recommendation?