[Backport to 3.4] Use plus tunings for lookahead scan more widely (#9813)

[bernhardmgruber]

Fixes: #9811

(cherry picked from commit 22e4e68)

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 9262d032-79ce-47a2-9ca8-eaf4cc4d02b9

📥 Commits

Reviewing files that changed from the base of the PR and between 305c95b and 03600bc.

📒 Files selected for processing (1)

• cub/cub/device/dispatch/tuning/tuning_scan.cuh

---

📝 Walkthrough

Summary by CodeRabbit

• Performance Improvements
  • Expanded warpspeed scan tuning support to additional scan operations when the accumulator meets the required compatibility conditions.
  • Enabled more input value sizes to use optimized SM10+ scan policies.

Walkthrough

The SM10+ warpspeed tuning gate now depends only on accumulator primitiveness or trivial copy constructibility, allowing matching policies for non-plus scan operations.

Changes

Warpspeed scan tuning

Layer / File(s)	Summary
Broaden warpspeed policy selection cub/cub/device/dispatch/tuning/tuning_scan.cuh	The SM10+ warpspeed policy switch no longer requires operation_t to be plus; the accumulator constraint remains enforced.

Assessment against linked issues

Objective	Addressed	Explanation
Restore scan_exclusive_custom performance on Blackwell [#9811]	❓	The tuning gate now covers non-plus operations, but no benchmark or regression-test evidence is provided to confirm the reported performance regression is resolved.

Suggested labels: backport branch/3.4.x

Suggested reviewers: elstehle, gonidelis, miscco

---

_{Comment @coderabbitai help to get the list of available commands.}

[bernhardmgruber]

pre-commit.ci autofix

[bernhardmgruber]

/ok to test a72fe92

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 1h 13m: Pass: 100%/283 | Total: 1d 21h | Max: 43m 26s | Hits: 100%/196179

See results here.