fix(kllama): stop leaving packed-quant matmul scalar-only on iOS/macOS

[michalharakal]

Summary

PlatformNativeKernels.ios.kt and .macos.kt were no-op stubs, both citing
SKaiNET#959 ("skainet-backend-native-cpu doesn't publish iosArm64/
iosSimulatorArm64/macosArm64 yet") as the reason. #959 shipped — the
engine has published real native-cinterop NEON kernels for those targets,
with runtime FEAT_DotProd dispatch, since then. linuxMain's actual already
calls installNativeKernels() to compensate for Kotlin/Native having no
ServiceLoader; the Apple actuals never got the same fix, so any app going
through kllama's registerPlatformBackends() facade has been silently
running packed-quant matmul on the scalar floor on iOS and macOS despite the
kernels being built and shipped.

Fix

Mirror linuxMain's actual on both files. kllama's nativeMain source set
already depends on skainet-backend-native-cpu and both iosMain/macosMain
dependsOn(nativeMain), so this is a code-only fix — no dependency wiring
needed.

Found while scoping out iOS support for the KernelRace sample in
SKaiNET-examples — this needed to land first, otherwise a new iOS target
would build and run correctly but silently stay scalar-only.

Test plan

• compileKotlinJvm, compileKotlinLinuxX64, compileTestKotlinLinuxX64, compileAndroidMain, compileKotlinJs all pass
• Apple targets (iosArm64/iosSimulatorArm64/macosArm64) — not verified, Kotlin/Native requires a macOS host to compile Apple targets at all, which this environment doesn't have. Verified by code review and cross-target compilation only.

[michalharakal]

Real-hardware validation of this fix (build + runtime kernel-tier check + before/after numbers) is tracked separately in #316, since it needs an actual Mac — not blocking this PR.

[michalharakal]

Validated on real Apple hardware (this needed a Mac, per the note above) and found + fixed a real bug: the original fix couldn't actually compile on any Apple target.

installNativeKernels() was unresolved on iosArm64/iosSimulatorArm64/macosArm64 — turns out implementation(libs.skainet.backend.nativeCpu) was still scoped to linuxMain's own dependencies block, not nativeMain, despite the commit message's claim. The engine side (#298/#959) is genuinely done — skainet-backend-native-cpu publishes iosArm64/iosSimulatorArm64/macosArm64 klibs now — but this module's build.gradle.kts never moved off the pre-#298 linux-only wiring. Pushed 9bb3707 moving the dependency up to nativeMain so all three Apple/Linux siblings inherit it.

After that fix, verified on this Mac:

• compileKotlinIosArm64 / compileKotlinIosSimulatorArm64 / compileKotlinMacosArm64 all succeed
• macosArm64Test and iosSimulatorArm64Test both green
• Runtime check: KernelRegistry.availableNames() reports [native-cinterop] on both macosArm64 and iosSimulatorArm64 after installPlatformNativeKernels() — not scalar
• compileKotlinJvm/LinuxX64/Js + compileTestKotlinLinuxX64 still pass (re-confirmed unaffected)

This covers the compile + runtime-registration checkboxes in #316. Real-device before/after throughput numbers still need actual hardware time, not just a compile+test pass — leaving that checkbox open there.