feat(inkling): reference configs, data pipeline, launcher, and v2 results

[Boreas618]

Item 15, rebuilt against main.

• configs: Small v1 (block 15) + YaRN, Small v2 (block 7, 6L all-full-attention, target-matched GQA, 8 uniform taps) stage-1 + YaRN-128/1M stage-2 — renamed to the existing upstream configs/inkling-dspark.json convention (no nda- prefix). The 947B-era gqa8/gqa16 variants are omitted (main already carries upstream's inkling-dspark.json); say the word if you want them too.
• data: regen corpus assembly with contract-keyed dedup, both stage-2 corpus builders, regen sampler.
• docs: v2 results — temp-0 9-task mean 3.936 vs v1 3.348 (+17.6%), flat long-context holdout (3.59/3.56/3.53 across 8K→32K+), colocated-topology memory notes.

The colocated launcher moved out of this PR — it belongs with the colocated trainer stack (#4), whose main-based form is pending your call.

🤖 Generated with Claude Code

[Boreas618]

Cross-reference: upstream sgl-project#734 and #735.

• Support Inkling managed capture sgl-project/SpecForge#734 adds a production-shaped disaggregated Inkling DSpark config + 2-node launcher (examples/configs/inkling-dspark-disaggregated.yaml) on upstream's new architecture — the topology twin of the colocated launcher here. It also makes --disable-radix-cache configurable (Inkling needs the unified radix tree), which our colocated engine handles via its own flags.
• feat: add Kimi K3 DSpark disaggregated reproduction sgl-project/SpecForge#735 (Kimi K3 disagg reproduction) carries 64K capture-kernel guards (int64 token offsets, scale-stable residual scoring) that live in the K3 SGLang patch — not applicable to our nda_sgl capture path (64K validated clean, probes green), but the failure class is worth knowing if K3-scale targets ever land on this lineage.

When upstreaming, these configs and the disaggregated ones should present as two topologies of one Inkling recipe.