A Differential Manifold Perspective and Universality Analysis of Continuous Attractors in Artificial Neural Networks