HIP: release cancelled prompt chunks at layer boundaries

[ThinkOffApp]

Cancelling a long single-GPU HIP Qwen prompt left the current 512/1,024-row chunk running through every layer: the lane's cancellation hook was not passed to the engine, and the engine's old 2,048-row polling threshold exceeded those chunks. The next short request therefore waited for the cancelled work.

Pass the stream cancellation hook to both prompt-step paths and poll at layer boundaries for single-rank passes of at least 512 rows. Shorter passes retain between-chunk cancellation. Tensor-parallel groups keep their existing behavior so ranks cannot diverge at a collective.

Fresh gfx1151/Qwen3.8-27B A/B results, three trials per build:

Measurement	Public main 76aafcf2	Fixed runtime 5429554e
Idle short request	1.899 s	1.913 s
Disconnect to engine idle metric	4.720–4.744 s	0.032–0.042 s
Next short request	6.532–6.548 s	1.902–1.907 s
Release/short-latency gate	3/3 fail	3/3 pass

All fourteen short replies have the same complete native token SHA-256. An identical uncancelled 14,975-token prompt also has identical native tokens; its prefill takes 79.021 s before and 78.867 s after (one timing control, not a broad performance claim). All twelve cancelled prompts fit the 16,384-token window and end as client-left with zero emitted tokens. Request lifetime is recorded separately from disconnect-to-idle time.

Validation:

• Both new CPU regressions fail with the old null hook/threshold and pass with the fix. The host test build passes 292/292 steps with cached results reused; its fresh results report 126 passes and two skips.
• The new serve_cancel_release.py gate checks admission, release latency, the following short request, and complete native token hashes; it retains the explicit historical 27B comparison behind --baseline-27b.
• The fixed server passes 13 exercised upstream served comparisons. Two drafting comparisons are unexercised because this checkpoint has no draft head.
• All 17 checkpoint files match revision 70ae7fac63274ff2eac54152031433374cb80f2f. Build/server commands, binary hashes, raw exchanges, original failing control and sanitization checksums are in the receipt directory.

Follow-up to the unresolved cancellation case in #994. This is scoped single-GPU 27B evidence, not a complete release qualification or a change to drafter support.

Current CI: source comments/docstrings were shortened to satisfy the lean rule (local comparison: zero new violations). Windows CI reports socket/dynamic-loader compile errors in unchanged platform paths, including speech.zig, cuda_rank_wire.zig and cuda_load_gpu.zig; the same errors are visible in an earlier unrelated run that predates this PR. Windows qualification is not claimed.