The Eigenlearning Framework: A Conservation Law Perspective on Kernel Regression and Wide Neural Networks