Learning Generalizable Shape Completion with SIM(3) Equivariance