Update base for Update on "update cuda delegate resource free pipeline for safety and segfault-free"

This diff survives `clear_all_tensors()` function and enable it during backend destroy stage. Furthermore, we defer the container handle deletion to OS to avoid potential segfault if there's more than one .so files.

Differential Revision: [D84135792](https://our.internmc.facebook.com/intern/diff/D84135792/)

[ghstack-poisoned]

Author: Gasoonjia

backends/cuda/runtime/cuda_backend.cpp

@@ -322,7 +322,7 @@ class ET_EXPERIMENTAL CudaBackend final
     if (handle->container_handle != nullptr) {
       AOTIRuntimeError delete_result =
           AOTInductorModelContainerDelete(handle->container_handle);
-      ET_CHECK_OR_LOG(
+      ET_CHECK_OR_LOG_ERROR(
           delete_result == Error::Ok,
           "Failed to delete AOTInductorModelContainer with error code %d",
           delete_result);
@@ -338,7 +338,7 @@ class ET_EXPERIMENTAL CudaBackend final
     if (!handle->so_path.empty()) {
       std::error_code remove_error;
       std::filesystem::remove(handle->so_path, remove_error);
-      ET_CHECK_OR_LOG(
+      ET_CHECK_OR_LOG_ERROR(
           !remove_error,
           "Failed to remove temporary shared library %s: %s",
           handle->so_path.c_str(),