onnruntime_shared_lib_test.exe passes/fails in non-deterministic fashion

@tianleiwu is already working on this.

Since Jun 16, 2026.

### Describe the issue

As the title says. I have a fresh build of 1.26 in RelWithDebInfo, with the CUDA EP, and onnruntime_shared_lib_test.exe fails about 40% of the time.

Looking at the definition of MyCustomKernelSecondInputOnCpu::Compute in custom_op_utils.cc, it seems to me that cudaMemcpyAsync shoud be used in line 101, instead of cudaMemcpy, to properly sequence the memcpy operation with the add operation. Otherwise, if the stream is non-blocking (as is the case in this test), the add kernel might run before memcpy has completed. Indeed, after changing the code that way and rebuilding, I don't observe failures anymore.

### To reproduce

Build 1.26 in RelWithDebInfo with the CUDA EP, then run `onnxruntime_shared_lib_test.exe --gtest_filter=CApiTest.custom_op_set_input_memory_type --gtest_repeat=20`.

### Urgency

_No response_

### Platform

Windows

### OS Version

11

### ONNX Runtime Installation

Built from Source

### ONNX Runtime Version or Commit ID

1.26

### ONNX Runtime API

C++

### Architecture

X64

### Execution Provider

CUDA

### Execution Provider Library Version

CUDA 12.4