Move Windows GPU training job to A10 (microsoft#19041)

### Description
1. Update sm to 86

### Motivation and Context
We have more A10 quota then T4 and Nvidia AXX could be  partitioned

tools/ci_build/github/azure-pipelines/win-gpu-ci-pipeline.yml

@@ -59,15 +59,14 @@ stages:
         BuildConfig: 'RelWithDebInfo'
         EnvSetupScript: setup_env_cuda.bat
         buildArch: x64
-        additionalBuildFlags: --enable_pybind --enable_training --use_cuda --cuda_home="$(Agent.TempDirectory)\v11.8" --skip_onnx_tests --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=75
+        additionalBuildFlags: --enable_pybind --enable_training --use_cuda --cuda_home="$(Agent.TempDirectory)\v11.8" --skip_onnx_tests --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=86
         msbuildPlatform: x64
         isX86: false
         job_name_suffix: x64_RelWithDebInfo
         RunOnnxRuntimeTests: ${{ parameters.RunOnnxRuntimeTests }}
         ORT_EP_NAME: CUDA
         WITH_CACHE: true
-        # Some unit tests crash on A10 GPUs. So this job still needs to use T4.
-        MachinePool: onnxruntime-Win2022-GPU-T4
+        MachinePool: onnxruntime-Win2022-GPU-A10
         isTraining: true
 
 - stage: dml