feat(clr): any-order overlap for oversubscribed shared HW queues.

[jaydeeppatel1111]

Motivation

When the number of HIP streams exceeds the hardware queue pool (GPU_MAX_HW_QUEUES, default 4), extra streams are recycled onto shared HW rings. Today every kernel keeps its head barrier bit set, so streams that merely share a ring serialize against each other even when they are logically independent — throughput collapses linearly with stream count. This adds an opt-in path that lets independent streams overlap on a shared ring, recovering most of the lost concurrency.

Technical Details

• New release flag DEBUG_HIP_SHARED_QUEUE_ANYORDER (default 0, opt-in). App must guarantee the sharing streams are independent (no cross-stream data hazard).
• In VirtualGPU::submitKernelInternal, clear the AQL head barrier bit for a stream's first packet since queue acquire on an oversubscribed ring. It reuses the existing any-order path (same bit-mask as hipExtAnyOrderLaunch).
• Gated by a conservative guard set so we only drop incidental same-ring serialization, never a real dependency:

1. last_write_index_ == kInvalidQueueIndex (stream's first packet since acquire),
2. eventWaitList().empty() (no explicit cross-stream dependency),
3. !cooperativeGroups()
4. !dedicated_queue_,
5. SharedHwQueueRefCount(queue, priority) > 1 (ring actually shared).

• Device::SharedHwQueueRefCount returns the pool share count for a HW queue, searching only the stream's priority pool (mirrors acquireQueue, honors DEBUG_HIP_IGNORE_STREAM_PRIORITY).
• Graph-capture exclusion (!isGraphCapture): graph replay encodes producer→consumer ordering as same-HW-queue in-order execution with an empty waitlist; clearing the barrier there would let a consumer overlap its producer. This guard keeps the graph executor's in-order invariant intact.

Issue Tracking

JIRA ID : AIRUNTIME-74

Test Plan

New correctness test unit/stream/hipSharedQueueAnyOrderOverlap.cc: independent oversubscribed streams produce correct results; explicit hipStreamWaitEvent dependencies are honored (must pass flag ON and OFF).
New perf test performance/scenarios/stream/hipPerfSharedQueueAnyOrderOverlap.cc: sweeps 2–32 streams, reports µs/iter.

Test Result

Full suite parity: ON and OFF on NV48
Performance (Performance_hipPerfSharedQueueAnyOrderOverlap, µs/iter) with streams 8, 16, 32 shows speedup 1.9x, 3.7x, 7.2x respectively.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/rocm-systems/blob/develop/CONTRIBUTING.md.

[therock-pr-bot]

❌ PR Check — Action Required

Check	Status	Details
🌿 Branch Name	✅ Pass	—
📝 PR Title/Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	❌ Fail	Error: Source/code files changed without an accompanying unit test. Expected: add at least one test file named like test_<name>.py / test_<name>.cpp (or <name>_test.*). Current: code file(s) changed: projects/clr/rocclr/device/rocm/rocdevice.cpp, projects/clr/rocclr/device/rocm/rocdevice.hpp, projects/clr/rocclr/device/rocm/rocvirtual.cpp, projects/clr/rocclr/utils/flags.hpp, projects/hip-tests/catch/performance/scenarios/stream/hipPerfSharedQueueAnyOrderOverlap.cc (+1 more); no test file found
🔎 pre-commit	⏳ Pending	⏳ Still running…
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—

⚠️ 1 policy check(s) failed. Please address the issues above before this PR can be Reviewed.

🚫 Please fix the failed policies

• ❌ Unit Test

The Not ready to Review label was added to this PR. Once all policies pass, the label is removed automatically.

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

[therock-pr-bot]

🚫 Please fix the failed policies before requesting reviews.

The following policy checks failed:

• ❌ Unit Test

The Not ready to Review label has been added to this PR.
Once all policies pass, the label will be removed automatically.