Self-distillation with Online Diffusion on Batch Manifolds Improves Deep Metric Learning