A Theory of Diversity for Random Matrices with Applications to In-Context Learning of Schrödinger Equations