Import Error in Windows debug config with eager mode

Describe the bug
When importing onnxruntime built from source for Windows in debug config with eager mode, we get a DLL load error for the module onnxruntime_pybind11_state.

Urgency
Blocking testing in windows debug

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Windows 11
• ONNX Runtime installed from (source or binary): source
• ONNX Runtime version: 1.12.0
• Python version: 3.9.13
• Visual Studio version (if applicable): NA
• GCC/Compiler version (if compiling from source): Clang 14
• CUDA/cuDNN version: NA
• GPU model and memory: NA

To Reproduce
Build from source:

1. Create and activate a venv
2. Install prereqs: pip install setuptools wheel numpy typing_extensions torch
3. Build for windows in debug mode: python ..\onnxruntime\tools\ci_build\build.py --config Debug --build_dir .\build --build_eager_mode --enable_training --build_wheel --parallel

The build stage itself will succeed, but the python tests will fail with a similar DLL error.

To repro once build succeeds:

1. Setup the python wheel (from Debug output dir): python ..\..\..\..\onnxruntime\setup.py bdist_wheel --enable_training
2. pip install the wheel: pip install "dist\onnxruntime_training-1.12.0+cpu-cp39-cp39-win_amd64.whl"
3. Try importing onnxruntime in interactive python shell

>>> import torch
>>> import onnxruntime

Expected behavior
Should be able to successfully import onnxruntime when built with Windows debug configuration and eager mode enabled

Screenshots
Error during python test stage of build:

Error during import after installing wheel:

Additional context
Both stack traces suggest that the error is specifically related to the onnxruntime_pybind11_state module.