Leveraging SE(3) Equivariance for Self-Supervised Category-Level Object Pose Estimation