A linear-attention layer that maintains a fixed-size recurrent state updated by gated projections, used in hybrid LLMs.