Upgrade llama.cpp Java binding to 4.2.0 and adapt adapters and loader to new API

[jsilvanus]

Motivation

• Unify the JVM and Android engines on the current de.kherud:llama:4.2.0 binding to ensure consistent GGUF compatibility and a request-oriented streaming API.
• Adapt the inference adapters to the newer llama binding API shapes so local inference and streaming work correctly with the updated native core.
• Improve streaming behavior so tokens are emitted lazily as they are generated and make model identification clearer via modelVersion.
• Avoid exposing a nullable artifact path on the public loader API by consolidating shared logic into a private implementation.

Description

• Bumped the llama binding dependency to de.kherud:llama:4.2.0 in engine/modelruntime and engine/androidapp and added clarifying comments about binding selection in the Android Gradle file.
• Updated LlamaCppAdapter to the new parameters and constructor shapes by using ModelParameters().setModel(...).setCtxSize(...).setThreads(...).setThreadsBatch(...).setBatchSize(...).setUbatchSize(...).setGpuLayers(...) and calling LlamaModel(params) instead of the old constructor, and updated modelVersion to java-llama.cpp-4.2.0.
• Changed generation code in both LlamaCppAdapter and AndroidLlamaCppAdapter to use InferenceParameters(prompt) with setNPredict, setTemperature, setTopP, and setTopK, and iterate model.generate(parameters) to stream token deltas as they arrive.
• Refactored ModelLoader to introduce a private loadInternal implementation that accepts a nullable artifactPath and exposed two public overloads for normal id resolution and exact artifact loading.
• Adjusted tests in GgufRot13FixtureTest to construct ModelParameters and InferenceParameters using the new API shapes.

Testing

• Ran ./gradlew :engine:modelruntime:test, which executed the JVM unit tests including GgufRot13FixtureTest, and the suite completed successfully (native inference checks are skipped when the platform lacks a native binary).
• Ran a full build with ./gradlew build to verify compilation across modules, and the build completed successfully.

---

Codex Task