Equivariant Adaptation of Large Pretrained Models