cleanup(gpu): delete GPU stubs and migrate native benchmark to DSL

[michalharakal]

Summary

There is no real GPU support in this repo — the GPU code paths in
`:llm-runtime:kllama` and the native benchmark engine are placeholders
that have never run on GPU and always fall back to CPU. This PR deletes
them and migrates the native benchmark to the DSL path, mirroring the
JVM cleanup from #127.

• Deleted: `GpuAttentionBackend.kt`, `GpuTensorBridge.kt` from kllama commonMain
• Stripped from kllama backend expect/actual (linux/macos/ios): `createGpuTensorBridge` and `createGraphAccelerator` (the latter was unused dead code — `GraphAccelerator` interface and the JVM `FusedQKVAccelerator` impl are CPU-side and stay)
• Stripped from llm-performance macosMain: `createMetalContext`, `createMlxContext`, `createGpuBridge` (only `availableNativeBackends` remains)
• Rewrote `NativeBenchmarkEngine`: dropped `GpuNativeLlamaAdapter` and the Metal/MLX scenario adapters; renamed scenario `native-backend-throughput` → `native-cpu-throughput`; migrated CPU adapter from legacy `LlamaRuntime` + `LlamaIngestion` to the DSL path (`DecoderGgufWeightLoader` + `LlamaNetworkLoader.fromWeights` + `OptimizedLLMRuntime` DIRECT)
• Updated `AttentionBackend` kdoc to drop the `GpuAttentionBackend` reference

Net: +73 / −479 across 9 files.

Test plan

• `./gradlew :llm-runtime:kllama:compileKotlinJvm :llm-runtime:kllama:compileKotlinLinuxX64 :llm-runtime:kllama:compileKotlinMacosArm64 :llm-runtime:kllama:compileKotlinWasmJs` — all green
• `./gradlew :llm-performance:compileKotlinJvm :llm-performance:compileKotlinMacosArm64 :llm-performance:compileKotlinWasmJs` — all green
• `./gradlew :llm-runtime:kllama:jvmTest :llm-performance:jvmTest` — pass
• (optional) Run `bash tests/smoke/smoke-test.sh` to confirm CLI smoke harness still passes

🤖 Generated with Claude Code