Supervised contrastive learning from weakly-labeled audio segments for musical version matching