On the Effect of Data-Augmentation on Local Embedding Properties in the Contrastive Learning of Music Audio Representations