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