Fix Qwen3 decode all-reduce perf validation

[chhwang]

Fix Qwen3 decode all-reduce perf validation.

Summary

• Add codegen external-buffer OFFSET handling.
• Copy packet all-reduce inputs into registered memory before mscclpp packet reads.
• Add C++ regression coverage for codegen and packet all-reduce setup.
• Add D2H-safe Qwen3 all-reduce equivalence tests.
• Rewrite the Qwen3 all-reduce bench to avoid torch GPU work while ARK's persistent runtime is launched.

Perf comparison

Measured on 8×A100, fp16, torch-free bench with host wall-clock timing and rt.barrier().

TP	Shape	Bytes	ARK	mscclpp ceiling	Ratio
2	decode (1,4096)	8 KB	5.31 µs	11.7 µs	0.45×
8	decode (1,4096)	8 KB	6.75 µs	11.7 µs	0.58×
2	prefill (2048,4096)	16 MB	990 µs	188 µs	5.27×
8	prefill (2048,4096)	16 MB	1046 µs	188 µs	5.56×

Decode passes the comm perf gate. Prefill does not pass and remains tracked by Q7P.

Benchmark rule

ARK's persistent runtime owns the GPU while launched. Do not issue torch GPU work (torch.cuda.synchronize, CUDA events, torch.allclose) during the benchmark. The bench uses ARK-safe host timing plus barriers and D2H-only correctness checks.

Validation

• Non-ROCm CI passed: lint, CodeQL, CUDA UnitTest, codecov, CLA.
• ROCm was ignored for this PR only.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 85.85%. Comparing base (c619d64) to head (803cbea).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #270      +/-   ##
==========================================
+ Coverage   85.78%   85.85%   +0.06%     
==========================================
  Files         129      129              
  Lines        6495     6498       +3     
==========================================
+ Hits         5572     5579       +7     
+ Misses        923      919       -4     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.