Latent Attention for Linear Time Transformers