Optimizing Likelihood-free Inference using Self-supervised Neural Symmetry Embeddings