Dual-granularity Sinkhorn Distillation for Enhanced Learning from Long-tailed Noisy Data