I want to wrap an existing torch.autograd.Function without modifying the internal to a custom operator, but the torch.library.register_autograd api requires backward and setup_context functions separately. What is the proper way for turning a torch.autograd.Function into custom_ops?