Rotational Equilibrium: How Weight Decay Balances Learning Across Neural Networks