SEM: Reinforcement Learning for Search-Efficient Large Language Models