[KDA] Default safe-gate Blackwell prefill to FlashInfer

[kpham-sgl]

Follow up to #41400: safe-gate KDA models now select FlashInfer prefill by default on SM100/SM103, including Kimi-K3 and GLM-5.3-Flash. The choice is resolved once per model runner through the existing linear-attention backend resolver.

Explicit prefill flags take precedence. Unbounded gates, other devices, deterministic inference, two-batch overlap, full prefill CUDA graphs, and explicitly selected non-Triton base backends keep their existing selection. Decode and verify selection are unchanged.

The CPU policy test covers model/hardware eligibility, operator overrides, and phase isolation. Its fixture publishes the resolved CUDA graph configuration, fixing the fixture failure from the earlier default-enablement attempt.

Validation: repository pre-commit checks, Python syntax checks, and git diff --check passed. Registered CPU, FlashInfer KDA kernel, Kimi-K3 B300, and GLM-5.3-Flash B200 suites will be requested through /rerun-test on this PR; these CI results are pending.

Prior opt-in validation and limitations are recorded in #41400 (comment). GLM accuracy/speed, Kimi Balanced/MegaMoE accuracy, and Kimi MMMU-Pro passed locally. The Kimi single-request speed gate failed for both prefill backends on the two-node GB300 setup, and the single-prompt acceptance difference remains unexplained. CI should validate the default on the registered B200/B300 recipes before merge.

---

CI States

Latest PR Test (Base): ❌ Run #37864294127
Latest PR Test (Extra): ❌ Run #37864293813
Latest PR Test (AMD ROCm 10): ❌ Run #37864294101

[kpham-sgl]

/rerun-test test/registered/e2e/models/test_kimi_k3_b300.py test/registered/e2e/models/test_kimi_k3_b300_low_latency.py test/registered/e2e/models/test_glm53_flash_b200.py test/registered/attention/test_kda_prefill_flashinfer.py test/registered/unit/layers/attention/test_linear_attn_config.py

[github-actions]

🚀 8-gpu-b300 (2 tests): ✅ View workflow run

cd test/ && python3 registered/e2e/models/test_kimi_k3_b300.py
cd test/ && python3 registered/e2e/models/test_kimi_k3_b300_low_latency.py

🚀 4-gpu-b200 (2 tests): ✅ View workflow run

cd test/ && python3 registered/e2e/models/test_glm53_flash_b200.py
cd test/ && python3 registered/attention/test_kda_prefill_flashinfer.py

🚀 ubuntu-latest (1 test): ✅ View workflow run

cd test/ && python3 registered/unit/layers/attention/test_linear_attn_config.py

[nvpohanh]

cc @leejnau