Add DSpark block drafter support

[Tianlei Wu (tianleiwu)]

Description

Adds end-to-end DSpark speculative decoding support for Qwen3.8 models. The implementation
reuses the DFlash2 block-drafter lattice runtime introduced by #2460, while adding DSpark
graph export, Markov scoring, shared target weights, and cache accounting for both full and
sliding-window draft models.

This PR is intentionally stacked on #2460 and should be reviewed after that PR.

Summary of Changes

DSpark model export

File	Change
src/python/py/models/builders/dspark.py	Builds the BF16 DSpark graph with dual-source non-causal attention, YaRN RoPE, target embedding and LM-head reuse, top-k candidates, and FP32 Markov lattice scores.
src/python/py/models/builders/qwen.py	Adds DSpark export options, validates target hidden-state taps, saves the auxiliary graph, shares external initializers, and emits model.dspark configuration.

Runtime integration

File	Change
src/config.cpp	Parses model.dspark through the common block-drafter configuration.
src/dflash2_drafter.h	Exposes cache sizing helpers used by engine pool planning.
src/dflash2_drafter.cpp	Accepts DSpark's all-predicting block geometry and supports fixed-ring or full-context cache allocation according to sliding_window.
src/engine/engine.cpp	Budgets full-attention drafter KV bytes with the main paged cache and retains fixed per-request rings for windowed drafters.

Testing

• cmake --build build/cuda130/Release --parallel
• ./build/cuda130/Release/engine_unit_tests
• python3 -m py_compile src/python/py/models/builders/dspark.py src/python/py/models/builders/qwen.py
• Standalone graph validation: 7/7 top-1 candidates matched; 109/112 top-k overlap.
• Greedy losslessness: 0 mismatches across 2,042 generated tokens.
• Eight-prompt acceptance test: 3.586 tokens per verification step over 1,862 steps.
• H200 benchmark matrix: 2.09x to 5.08x decode throughput over the same target graph without a drafter across context lengths 512-8192 and batch sizes 1-16.
• MMLU-Pro 800 samples: 82.62% with DSpark versus 83.00% without drafting (p=0.508).
• GPQA Diamond: 76.26% with DSpark versus 72.73% without drafting (p=0.143).

Motivation and Context

DSpark extends the DFlash-style parallel block drafter with a low-rank Markov head. Both
models expose the same candidate lattice to the runtime, so sharing the proposal and
verification path avoids a second speculative-decoding implementation.

The evaluated Qwen3.8 checkpoint uses full attention and therefore grows its draft KV cache
with resident context. DSpark is not inherently limited to full attention: positive
sliding_window configurations use the existing fixed-ring path.

Checklist

• Tests added/updated through the shared block-drafter test coverage
• No breaking changes
• Export and runtime behavior validated end to end

[Tianlei Wu (tianleiwu)]

Superseded by #2497, which reuses the compact #2496 block-drafter runtime for DSpark without duplicated builder changes, dense recurrent checkpoints, or an ONNX Runtime dependency bump.