perf(qwen): reduce generation overhead

[chaofengw-nv]

Background

qwen3-4b-instruct-2507 preserved exact FP16 output but measured 39.7% slower than its reference on Thor X. This change removes repeated CUDA-graph setup from the generation path and keeps the release comparison precision aligned.

Exit Criteria

• Preserve exact-token Accuracy behavior on Thor X and GB300.
• Bring the Thor X release workload back within the existing 5% performance policy without weakening the gate.
• Keep the fixed native KV-cache boundary valid for full-capacity prompts.

Implementation

• Enable CUDA graphs for the Qwen3-4B release workload and compare against an FP16 reference.
• Skip decoder graph priming once the graph is already captured.
• Account for the first generated token coming from prefill so fixed KV capacity only reserves rows consumed by later decoder steps.
• Add source and native-contract regressions for graph priming, release policy, and KV-cache capacity boundaries.

Validation

• python3 -m pytest -q tests/builder/test_qwen_runtime_performance_contract.py tests/e2e/models/qwen/test_qwen_performance_contract.py: 4 passed.
• python3 -m tools.community_ci source-quality --base github/main: passed all complexity, lint, architecture, and source-quality checks; 158 tests passed.
• test_qwen_native_kv_cache native executable: passed on GB300 at exact head 500ca807556eaa5a5b8d59717af559dfe41889dc.
• GB300 model checks at the exact head: Accuracy passed 20/20 with exact agreement and 0 failed samples; Perf passed green with 9.34 ms TRTMC p50 versus 71.31 ms reference p50.
• Thor X model checks at the exact head: Accuracy passed 20/20 with exact agreement and 0 failed samples; Perf passed green with 102.80 ms TRTMC p50 versus 109.88 ms reference p50. Both 10-sample timing series were stable and output tokens matched exactly.
• A broader local Qwen collection reached 298 passed and 4 skipped; 11 GPU/E2E cases were not runnable in the local no-CUDA/read-only-storage environment and were covered by the target-hardware checks above.

Notes For Future Readers

• Review the runtime changes first, then the Qwen-specific release profile, followed by the shared native KV-cache contract tests.
• The existing 5% performance policy is unchanged.

Closes #901