Symmetric Reinforcement Learning Loss for Robust Learning on Diverse Tasks and Model Scales