RFE: Add deterministic option for reduce, scan, etc.

[ekelsen]

The default work sharing strategy on sm_35+ isn't deterministic by default. Would be great to have the option to use the even_share strategy to get determinism even if it's a little slower.

[dumerrill]

Commit f558c6d resolves this issue for all variants of device-wide reduction. (Block-wide reduction, warp-wide reduction, and all variants of prefix scan have always been deterministic for pseudo-associative operations.)

Note that all CUB device-wide algorithms are "run-to-run" deterministic for pseudo-associative reduction operators (like floating point addition), i.e., their results are consistent across multiple runs on the same device. Floating point reduction may produce slightly different results on different devices having different numbers of SMs or different compute capabilities (where CUB might use different tile sizes.)

[lilohuang]

@dumerrill comment might be incorrect in 2022 especially for some device-wide operations. see details in #441 and NVIDIA/thrust#1587 (comment)

CC @allisonvacanti @senior-zero

[alliepiper]

Closing after all, we'll let this old RFE rest. I've filed a new issue for the regressions at NVIDIA/cccl#886.