onnxruntime

Fix race condition in CUDA, ROCm, and TensorRT EP GetKernelRegistry() implementations.

#10200

Merged