hotfix for memory leak in aot autograd induced by saving tensors for backward (#95101) (#95477)
Workaround fix in AOTAutograd for https://github.com/pytorch/pytorch/issues/94990 (see the comments for more details / discussion)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/95101
Approved by: https://github.com/albanD