41. Align before Fuse: Vision and Language Representation Learning with Momentum Distillation