Reward Design with Program Graphs for Reinforcement Learning Guided Training of Large Language Models for Program Synthesis