<em>Simpler is Better:</em> Finding the Best Reward Function in Long Chain-of-Thought Reinforcement Learning for Small Language Models