Recurrent GDN/KDA for decoding phase

[learning-chip]

https://github.com/huawei-csl/megagdn-pto only does chunk GDN for prefill. sgl-kernel-npu/fused_sigmoid_gating_recurrent.py has Triton baseline for both GDN and KDA decoding.

This is an easy pure-vector memory-bound kernel, so not expecting much room for performance improvements. Just for feature completeness.