Interpreting Attention Layer Outputs with Sparse Autoencoders