[CUDA] Avoid repacking NVFP4 QMoE decode weights (#32325)
## Description
Make the NVFP4 QMoE decode GEMV consume the schema-native packed weights
and scales directly. This removes the decode-only weight repack and
activation-dtype scale bank while retaining the existing grouped-GEMM
paths for larger workloads.
## Summary of Changes
### Raw NVFP4 GEMV
| File | Change |
|------|--------|
| `onnxruntime/contrib_ops/cuda/llm/moe_gemm/moe_gemv_fp4.cu` | Add a
raw N-packed NVFP4 GEMV that decodes E2M1 weights and E4M3 scales
in-register, including fused SwiGLU. |
| `onnxruntime/contrib_ops/cuda/llm/moe_gemm/moe_gemv_fp4.h` | Extend
support checks and launch interfaces for the raw NVFP4 layout. |
| `onnxruntime/contrib_ops/cuda/moe/moe_quantization.cc` | Retain and
dispatch directly from raw NVFP4 initializers instead of building
decode-specific packed weights and combined scales. |
| `onnxruntime/contrib_ops/cuda/llm/moe_gemm/moe_kernels.cu` | Extend
the fused expert-map prologue to Qwen's `top_k=10` configuration. |
### Tests
- Add the Qwen Flash decode shape and verify that it selects the GEMV
route.
- Cover disabled prepacking and fallback behavior above the GEMV row
limit.
## Testing
- `python3 -m pytest -q
onnxruntime/test/python/transformers/test_qmoe_nvfp4_cuda.py`
- 32 passed on NVIDIA H200.
- CUDA provider target built successfully with FP4 QMoE enabled.
- `ruff check
onnxruntime/test/python/transformers/test_qmoe_nvfp4_cuda.py`
## Checklist
- [x] Tests added/updated
- [x] No breaking changes
- [x] Documentation not required
