onnxruntime

c98a0e01 - update CMAKE_CUDA_ARCHITECTURES for packaging pipelines (#25126)

update CMAKE_CUDA_ARCHITECTURES for packaging pipelines (#25126) * Change 90 to 90a. It is because FpA IntB Gemm uses accelerated features (like WGMMA, TMA and setmaxnreg). * Change 90 to `90a-real;90a-virtual`. It is because a recent change in [cuda_configuration.cmake](https://github.com/microsoft/onnxruntime/blob/main/cmake/external/cuda_configuration.cmake) will replace "90" to "90a-real", and we need explicitly add the virtual one `90a-virtual` to the list. Note that our pipelines are still using CUDA 12.2, so it is not able to add 100 and 120 to the list. We can add it when we upgrade to CUDA 13 in the future.