Benchmarking Potential Based Rewards for Learning Humanoid Locomotion