Add sm90 and sm1200 policies for segmented_sort

[gonidelis]

Fixes #6173 by applying sm80 tunings parameters to sm90 compilations (server) and sm86 tuning parameters to sm120 compilations (workstation)

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[gonidelis]

Posting some results in Google Sheets instead of laying them here for readability Reasons

[gonidelis]

B200 (SM100)

SM100 perf results spreadsheet

[gonidelis]

NVIDIA GeForce RTX 5090 (SM120)

SM120 perf results spreadsheet

[gonidelis]

NVIDIA H100 PCIe (SM90)

SM90 perf results spreadshseet

[gonidelis]

I did some thorough looking into the sm90 results and for power law distribution segments the only problematic workload is when OffsetT{ct}=I32 Elements{io}=2^22 Segments{io}=2^20 (highlighted rows). This regresses a lot but all the other workloads show significant speedup. The thing is that segment sizes are a runtime param so we cannot have control over it so when to choose the tunings. large and small regress significantly and in a different pattern. So maybe for the time being we should defer merging this.

[gonidelis]

Closing as it fixes some cases but regresses others. Read detailed explanation in the tracking issue.