What Affects Learned Equivariance in Deep Image Recognition Models?