Multi-Frame Self-Supervised Depth with Transformers