Contrastive Sparse Autoencoders for Interpreting Planning of Chess-Playing Agents