v1.3: fused multi-chunk backward (kda_chunked_fused, 2-5x faster training)

v1.2: training / backward support (kda_recurrent, kda_chunked + backward kernels)

v1.1: T-KDA-02 fix + fully model-agnostic (remove all consumer-model references)

v1.1: T-KDA-02 state-decay axis fix + parity update