Describe the bug
Our application builds a shared library, ort_wrapper, using ORT C API, use dlopen(), dlclose() on ort_wrapper to load / unload ORT dynamically. It works fine for EPs like CUDA and OpenVINO, but it results in segfault if TRT EP is used. Stack trace shows

Thread 1 "test" received signal SIGSEGV, Segmentation fault.
0x00007fd78091def6 in (anonymous namespace)::KernelRegistryAndStatus::~KernelRegistryAndStatus() () from /workspace/build/Release/libonnxruntime_providers_tensorrt.so
(gdb) bt
#0  0x00007fd78091def6 in (anonymous namespace)::KernelRegistryAndStatus::~KernelRegistryAndStatus() () from /workspace/build/Release/libonnxruntime_providers_tensorrt.so
#1  0x00007fd7aa72e6c5 in __cxa_finalize (d=0x7fd780d11580) at cxa_finalize.c:83
#2  0x00007fd780919b13 in __do_global_dtors_aux ()
   from /workspace/build/Release/libonnxruntime_providers_tensorrt.so
#3  0x00007ffd744459a0 in ?? ()
#4  0x00007fd7aacf5ccc in _dl_close_worker (map=<optimized out>, 
    force=<optimized out>) at dl-close.c:288
Backtrace stopped: frame did not save the PC

Urgency
High, we are aiming to update ORT to 1.5.1 for the release.

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Ubuntu 18.04
• ONNX Runtime installed from (source or binary): source
• ONNX Runtime version: 1.5.1
• Python version: 3.6
• Visual Studio version (if applicable):
• GCC/Compiler version (if compiling from source): gcc (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0
• CUDA/cuDNN version: 11.0.221 / 8.0.4.12
• GPU model and memory:

To Reproduce

• Obtain the Docker file, C++ example, and model (resnet) https://drive.google.com/drive/folders/19GLdI9aiEmJrsHYjaWj1RoZ9ckouzlG0?usp=sharing
• Build docker image docker image build -t ort_bug .
• Run image with GPU docker run --gpus=0 -it ort_bug:latest
• Run example with TRT EP ./example -b (remove -b to run on default EP), which will be terminated with Segmentation fault (core dumped)

Expected behavior
Program should exit normally