On Feature Learning in Neural Networks with Global Convergence Guarantees