cub device scan is not deterministic as described in the documentation

[zasdfgbnm]

Issue originally reported at pytorch/pytorch#75240
To repro, see: #453

[wence-]

FWIW, 7a8320c updates the docs to indicate that these operations are not deterministic run-to-run. But it looks like there hasn't been a release to build them since then?

[alliepiper]

Unfortunately this issue has been popping up a lot recently in several algorithms, scan and reduce variants in particular. It appears that this documentation has been incorrect for quite some time, as these algorithms are implemented using techniques like decoupled lookback that are completely non-deterministic.

I'm aware of the situation and would like to fix it, but currently lack the resources needed to address these problems. This situation should hopefully improve in the next few months and ideally we'll be able to provide deterministic versions of these algorithms at some point.

[alliepiper]

Wrote up a summary of the issue and some proposed solutions in NVIDIA/cccl#886. Feedback is welcome -- closing this issue to consolidate discussion to the new issue.