[Bugfix][SM70] Gate DFlash2 verifier by native capability

[yangzhuxinyzx]

Purpose

Prevent a source-overlay deployment from routing q16 DFlash2 verification into an older Flash-V100 native extension that only supports q8. The production incident occurred on a 51,088-token cached request with a 16-token continuation: the native TORCH_CHECK killed EngineCore and left the systemd service falsely active with orphaned TP workers.

Base SHA: e315dd47096524ed8a9fd43770f925a188a59d23.
Current head SHA: f1f5741cda.

The native extension now exposes its grouped-verifier max-query capability. Python defaults legacy binaries without that entry point to q8, while newly rebuilt q16 binaries report 16. The normal DFlash2 q8 verifier remains on the grouped fast path; unsupported q16 work falls back to the exact paged-decode path.

Test Plan

• exercise q8 and q16 grouped routing with a simulated q16-capable extension;
• exercise q16 fallback with a simulated legacy q8 extension and assert the grouped operator is never called;
• load an existing legacy Flash-V100 binary and verify the compatibility fallback reports max-q 8;
• run changed-file pre-commit and Python byte compilation;
• deploy to the four-V100 service and replay health, normal q8 decode, a 16-token final prefill chunk, prefix cache, and tool calling.

Test Result

• policy tests: 5 passed, 112 deselected;
• legacy binary capability probe: native entry absent, reported max-q 8;
• changed-file pre-commit, Markdown lint, Python compileall, and git diff --check: passed;
• exact incident shape: 49,084 prompt tokens = 12 x 4,089 + q16; completed in 13.288 s without an engine failure;
• repeated 49,084-token prefix: 1.715 s, confirming prefix-cache health;
• tool smoke: valid get_weather({"city":"北京"}) call;
• historical web prompt: 512 output tokens, 206.06 token/s streaming decode, 17.463 ms/engine round, 3.599 emitted tokens/round;
• service health: active/running, public and loopback health checks pass, no post-restart traceback or grouped-verifier shape error.