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

@@ -337,11 +337,12 @@ class ET_EXPERIMENTAL CudaBackend final
     // Remove the temporary shared library file
     if (!handle->so_path.empty()) {
       std::error_code remove_error;
+      std::filesystem::remove(handle->so_path, remove_error);
       ET_CHECK_OR_LOG(
           !remove_error,
           "Failed to remove temporary shared library %s: %s",
           handle->so_path.c_str(),
-          remove_error.message().c_str())
+          remove_error.message().c_str());
     }
 
     delete handle;