Learning Deep Optimal Embeddings with Sinkhorn Divergences