perf(inference): chunked-parallel GatedDeltaNet scan (race-fixed, gated)

[ohdearquant]

chunked-parallel GatedDeltaNet scan (race-fixed, gated)

Stacked on #189. Adds a chunked-parallel (Yang 2024, C=32) GatedDeltaNet scan for
the 18 GDN layers, replacing the serial recurrence in prefill. Env-gated
(LATTICE_GDN_CHUNKED); the serial path remains the default.

Result — interleaved same-process A/B (Qwen3.5-0.8B, Apple Silicon, 18 GDN layers)

Context (tok)	serial tok/s	chunked tok/s	Δ median	chunked CV
289	459.7	608.5	+32.4%	3.2%
529	413.9	585.6	+41.5%	1.8%
1009	390.7	549.2	+40.6%	1.5%

Correctness — the interesting part

The chunked path first failed correctness (nondeterministic state up to 0.45). The fix:

1. A standalone fp32 CPU oracle of the chunked algebra (W/U forward-sub, R = U − γ·S₀ᵀW,
   S_end, QKL output) matched the serial recurrence to 1e-7 — so the algebra was
   never wrong, disproving an earlier "state-algebra error" diagnosis.
2. Real bug: 4 missing threadgroup_barrier — cross-simdgroup WAR races on the
   shared reduction scratch (materialize ×2, solve ×2). A lagging simdgroup read the
   next reduction's partial → corrupted normalization.
3. Localized headless via a gdn_chunked_race_localization_probe test — no GPU capture
   needed.

Gates (post-fix, all green)

gate	threshold	result
B-vs-B self-consistency	< 1e-3	0.000000
chunked-vs-chunked state	< 1e-3	0.000000 (15+ runs)
parity (logits)	argmax stable	6.3e-5, argmax stable
chunked-vs-serial state	magnitude-aware	2.2e-4, 10/10

Notes

• 0.8B shape only (kd=vd=128), prefill only, batch=1. Default-flip is a follow-up once
  it has mileage.
• No new crates; no library unwrap().