S2SD: Simultaneous Similarity-based Self-Distillation for Deep Metric Learning