[CUDA] Qwen3.6-35B-A3B Throughput Optimization

[tianleiwu]

This is used to track the progress of Qwen3.6-35B-A3B Throughput Optimization.

Related PRs:
For Qwen 3.6:

• olive-recipes
  Add Qwen3.6-35B-A3B MoE VLM recipe (CUDA + CPU)

• onnxruntime-genai
  Fix CUDA QMoE INT4 export for Qwen3.5/3.6 MoE models
  Qwen3.6 MTP onnxruntime-genai#2218

• cuda op / kernels:
  [CUDA] Optimize QMoE SoftmaxTopK router for small-batch decode #28980
  [CUDA] Add decode-optimized LinearAttention (GatedDeltaNet) kernels #28985
  [CUDA] Add decode (M=1) GEMV fast path to MatMul #28986 (Not needed if we have shared expert optimization like below)
  [CUDA] QMoE support shared experts #29028
  [CUDA] QMoE GEMV fast path for batch-1 decode #29038 (Need to extend to block quantization)
  Use user-supplied external initializer in place when already on the planned device #29013

Related issues:
AddExternalInitializers copies device (GPU) OrtValues per session instead of using them in place