On Privileged and Convergent Bases in Neural Network Representations