[cuda.compute]: Fix segmented sort selector race

[NaderAlAwar]

Description

This separates the race fix found by TSan in #9986 into it's own PR (and removes some redundant std::free calls)

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[NaderAlAwar]

/ok to test 33c0a35

[coderabbitai]

📝 Walkthrough

Summary by CodeRabbit

• Bug Fixes

  • Improved segmented sorting stability and reliability by managing selector operation state independently for each sort execution.
  • Updated serialized segmented-sort operations to restore correctly without retaining unnecessary runtime state.

• Performance

  • Reduced persistent memory usage associated with compiled selector operations.
  • Simplified cleanup during segmented-sort execution and deserialization.

Walkthrough

Changes

Segmented sort selector state

Layer / File(s)	Summary
Selector operation contract c/parallel/src/segmented_sort.cu	Generated selector operations now return state = nullptr; ownership and compilation wiring no longer manage persistent selector state.
Per-dispatch selector state c/parallel/src/segmented_sort.cu	Selector functors receive per-call state pointers, while dispatch creates stack-local large and small selector states and passes them into the kernel source.
Serialization and cleanup alignment c/parallel/src/segmented_sort.cu	Deserialization restores selector operations without state and updates failure and cleanup paths to release only selector code buffers.

Suggested reviewers: oleksandr-pavlyk

---

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

[coderabbitai]

🧹 Nitpick comments (1)

c/parallel/src/segmented_sort.cu (1)

269-320: 🩺 Stability & Availability | 🔵 Trivial

Confirm regression test coverage for the fixed race.

The PR comments note testing was requested for the commit fixing this race. No test changes are included in this file/batch. Given the C API path instructions call for coverage of public C parallel APIs, and this is specifically a concurrency fix (hardest class of bug to regress-test), it's worth confirming a test exercising concurrent cccl_device_segmented_sort builds/dispatches on a shared build result exists (e.g. under a ThreadSanitizer-enabled CI job), per the referenced .agent/skills/cccl-test/SKILL.md guidance for this file type. Happy to help draft such a test if none exists yet.

As per coding guidelines, "When writing, updating, reviewing, or validating tests, follow the applicable guidance in .agent/skills/cccl-test/SKILL.md" and per path instructions, "Focus on ... coverage for public C parallel APIs" for c/**/*.

Also applies to: 904-925

Sources: Coding guidelines, Path instructions

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 91076c03-f3f8-49c8-a1be-ea379b940949

📥 Commits

Reviewing files that changed from the base of the PR and between 73a1a4b and 33c0a35.

📒 Files selected for processing (1)

• c/parallel/src/segmented_sort.cu

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 2h 44m: Pass: 100%/20 | Total: 6h 03m | Max: 49m 14s | Hits: 98%/1515

See results here.