An Empirical Study on Clustering Pretrained Embeddings: Is Deep Strictly Better?