Revisiting Neural Networks for Continual Learning: An Architectural Perspective