[CUDA] GPT-OSS-20B Throughput Optimization

[tianleiwu]

This is used to track the progress of GPT-OSS-20B Throughput Optimization.

Related PRs:

• olive-recipes
  Add GPT-OSS 20B recipes olive-recipes#507

• onnxruntime-genai
  Update model builder for gpt-oss onnxruntime-genai#2234

• cuda kernel improvements:
  [CUDA] QMoE GEMV fast path for batch-1 decode #29038
  [CUDA] Optimize FlashDecode split planning for local-window GQA #29161
  [CUDA] Enable XQA decode for GroupQueryAttention with attention sink #29162
  [CUDA] Default QMoE GEMV fp16 accumulation for fp16 activations #29166
  [CUDA] Add sliding-window support to XQA decode #29177
  [CUDA]: Split-K2 QMoE SwiGLU GEMV kernel #29167 (Experiment)

• fusion
  [CUDA] Enable CUDA GQA QK-Norm and XQA decode #29186
  [CUDA] Fuse MoE router bias into MatMulNBits GEMV #29170
  QMoE router Fusion (Experiment)

• Experiemnts
  quantization recipes experiments

BatchSize=1, H200 GPU

Quantization	Size (GiB)	Prefill TPS	Decode TPS	MMLU
rtn, mixed, per-channel MoE, int4 lm_head	10.77	29403	382.0	0.8017
k_quant, mixed, per-channel MoE, int4 lm_head	10.79	29018	366.7	0.8128
k_quant, mixed, block-64 MoE, **int8 lm_head	11.61	24492	345.5	0.8170