Self-Supervised Representations Improve End-to-End Speech Translation