v1.5: exact chunked backward (kda_chunk_step_exact_bwd)

v1.4: forward optimizations (decode key-fold, batched decode, exact prefill, head-interleaved prefill)

v1.3: fused multi-chunk backward

v1.2: training / backward support

v1.1: T-KDA-02 state-decay axis fix + fully model-agnostic

v1.0: KDA NKI kernels for HF Transformers on AWS Trainium