REINFORCE++: A Simple and Efficient Approach for Aligning Large Language Models