`linalg_det_singular`: modify samples such that CUDA IMA dissapears. (#72585)
Summary:
Implicitly fixes https://github.com/pytorch/pytorch/issues/72203 and https://github.com/pytorch/pytorch/issues/72204.
The issues is coming from an incorrect use of `scatter` with wrong indices, see https://github.com/pytorch/pytorch/issues/72204#issuecomment-1034087199. I do not know what exactly calls to `scatter`, investigating...
Pull Request resolved: https://github.com/pytorch/pytorch/pull/72585
Reviewed By: cpuhrsch
Differential Revision: D34245279
Pulled By: anjali411
fbshipit-source-id: 460f030524f9228f2269eaee0a3a72e1978caeb4
(cherry picked from commit e48295716aa00078024d8b1d196358d6bc926b83)