perf(qwen35): accelerate Ampere prefill GDN

[dusterbloom]

Summary

• keep Q/K at their compact 16-head shape for fused Qwen3.5 Gated DeltaNet instead of materializing two 16-to-48 repeats
• select the existing grouped-column GDN kernel automatically for NVIDIA Ampere prefill launches of at least 512 tokens while retaining the classic decode/partial-chunk path
• preserve the existing force/disable diagnostics and the opt-in chunked fallback's required Q/K repeats
• add pure selector, graph-structure, and focused CUDA numerical parity tests

Dependency

This is stacked on #518 (perf(cuda): reduce MMQ stream-k overhead). The branch is current main (0e002364) plus #518 (506cd972) plus this commit (3c0be1f). Once #518 merges, this PR's visible diff should collapse to the GDN/QK change only.

Performance

Matched Q4_K_M / F16-F16 KV prefill on one RTX 3090, CUDA 12.6, one slot, prompt caches disabled, one generated token, default ubatch 512, and no GDN/ubatch environment overrides. Each value pools ten measured samples from an A-B-B-A run with one warmup and five measurements per phase.

Prompt depth	Lucebox stack	llama.cpp	Delta
1K	1341.47 tok/s	1278.72 tok/s	+4.91%
8K	1393.73 tok/s	1359.53 tok/s	+2.52%
16K	1332.63 tok/s	1306.86 tok/s	+1.97%

The deltas are approximately 7.0x, 7.3x, and 5.5x their combined standard errors. The exact same prompt payload and model artifact were used, but the runtimes report a two-token accounting difference (988/986, 8156/8154, and 16349/16347), so this is not presented as bit-identical tokenizer accounting. The raw result's AutoLuce content ID is measurement-792d2c28b337ed544dc8c1e3e3bd699c65d8c2b4f182813dfac272c1359fb047.

Correctness

• gdn_grouped_policy: verifies the 512-token Ampere threshold and override precedence
• qwen35_gdn_broadcast: verifies compact Q/K reaches fused GDN and the chunked fallback retains exactly two repeats
• gdn_grouped_cuda: compares forced classic and grouped output/final state on separate graphs with freshly restored state at 1, 221, 477, 512, and 768 tokens using the real 16-Q/K-head / 48-value-head shape; all NMSE values are approximately 1e-14, below the 1e-6 gate

Verification

cmake --build /tmp/lucebox-prefill-gdn-q4-build --target dflash_server test_gdn_grouped_policy test_qwen35_gdn_broadcast test_gdn_grouped_cuda -j4
ctest --test-dir /tmp/lucebox-prefill-gdn-q4-build -R '^(gdn_grouped_policy|qwen35_gdn_broadcast|gdn_grouped_cuda)$' --output-on-failure

All three tests pass on SM86.