Latent Positional Information is in the Self-Attention Variance of Transformer Language Models Without Positional Embeddings