Prevent module full_backward_hook from erroring in double backward (#88357)
Also clarifies documentation to say "execute if and only if gradients wrt outputs are computed" (previously, "execute every time gradients wrt inputs are computed")
See https://docs.google.com/document/d/1tFZKYdsSzRBJ7Di7SWt8X8fSg-E3eiUPwomMF10UyhM/edit for more details regarding the question: 'should module full_backward_hooks be called every time the gradients wrt module inputs are called, or should module full_backward_hooks only be called when the "backward for the module" have been computed?'
Fixes https://github.com/pytorch/pytorch/issues/88312
Pull Request resolved: https://github.com/pytorch/pytorch/pull/88357
Approved by: https://github.com/albanD
