build(llm-inference): wire native-cpu provider into qwen + llama jvmTest

[michalharakal]

Summary

Wires the priority-100 native (FFM) kernel provider into the jvmTest classpaths of llm-inference/qwen and llm-inference/llama so pipeline tests exercise the native Q4_K + FP32 kernels — 4–6× faster than Panama on Q4_K matmul, 1.5–1.8× on FP32 SGEMM (per upstream microbench numbers from SKaiNET PRs #572 and #575).

Pairs with upstream PR SKaiNET#576 which adds publishing config for skainet-backend-native-cpu. Composite build (includeBuild("../SKaiNET") in settings.gradle.kts:21) handles the substitution; no SKaiNET release needed for local dev.

Skips llm-inference/gemma deliberately. Gemma has its own unrelated stability issues today (chat-template / tool-calling gaps per workspace context); validating perf wins on a known-broken inference path would muddy the signal. Same one-line change applies once gemma stabilizes.

Changes

• gradle/libs.versions.toml — add skainet-backend-nativeCpu alongside the existing skainet-backend-cpu. Catalog key uses camelCase (nativeCpu) rather than dashes because Gradle's type-safe accessor generator chokes on native as a path segment (it's a soft-reserved Kotlin keyword); the underlying Maven coordinate stays kebab-case sk.ainet.core:skainet-backend-native-cpu.
• llm-inference/qwen/build.gradle.kts + llm-inference/llama/build.gradle.kts — add implementation(libs.skainet.backend.nativeCpu) to the jvmTest source set, parallel to the existing skainet.backend.cpu entry. JVM-only (FFM has no Native / JS / Wasm equivalent).

Wiring contract

DefaultCpuOpsJvm already calls KernelServiceLoader.installAll() lazily on first use, so any test that exercises matmul through ctx.ops automatically picks up the native provider when it's available. No runtime code changes anywhere — pure dependency-graph + auto-discovery via META-INF/services.

Test plan

• :llm-inference:qwen:jvmTest — passes; QwenDslPipelineTest 6/6, QwenConfigParserTest 2/2
• :llm-inference:llama:jvmTest — passes; LlamaDslPipelineTest 6/6, StateManagementTest 12/14 (2 pre-existing skips), LlamaWeightMapperTest + LlamaQuantDequantTest pass
• CI confirms green with the local SKaiNET → published 0.22.0-SNAPSHOT swap (this PR depends on upstream publishing landing)
• On hosts where the native lib doesn't load (sandbox, missing arch), KernelRegistry cleanly cascades to Panama priority-50 — the SKaiNET native-cpu module's own NativeFfmPipelineTest already exercises this fall-through path

🤖 Generated with Claude Code