Dont increment generation if forward of backward exists, and warning on deallocation of live tensors (#97168)
Refining the logic for when it is okay to ignore previously live outputs from cudagraphs. If there is a forward that has been invoked without invocation of the corresponding backwards, dont allow overwriting outputs.
Differential Revision: [D44228369](https://our.internmc.facebook.com/intern/diff/D44228369)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97168
Approved by: https://github.com/ezyang, https://github.com/jansel