[CUDA] Normalize the PER_CHANNEL K scale folded into Q for paged XQA - #32520

[CUDA] Normalize the PER_CHANNEL K scale folded into Q for paged XQA#32520

Tianlei Wu (tianleiwu) merged 2 commits into

tlwu/20260909/int4_kv_per_channelmicrosoft/onnxruntime:tlwu/20260909/int4_kv_per_channelfrom

tlwu/20260909/int4_kv_fold_scale_normmicrosoft/onnxruntime:tlwu/20260909/int4_kv_fold_scale_normCopy head branch name to clipboard