SPSD: Similarity-preserving self-distillation for video–text retrieval