Cross-Attention is All You Need: Adapting Pretrained Transformers for Machine Translation