ADFA-3105 - Wire debug llama asset bundling and add versioned dynamic AAR refresh for Local LLM

[jatezzz]

This PR wires the local LLM (llama.cpp) update through the full asset lifecycle so both debug and release users get the new runtime library consistently, including Brotli-based assets.

What changed

1) Debug build flow now bundles updated llama assets

Updated app/build.gradle.kts so debug variants explicitly depend on llama asset bundling and debug asset sync:

• assembleV8Debug depends on:
  • bundleV8LlamaAssets
  • assetsDownloadDebug (non-CI)
• assembleV7Debug depends on:
  • bundleV7LlamaAssets
  • assetsDownloadDebug (non-CI)

This matches release behavior and ensures debug builds also regenerate/pick up the latest llama assets.

2) Runtime refresh for installed llama AAR

Updated DynamicLibraryLoader to force-refresh llama.aar from bundled assets when version changes:

• Reintroduced ensureLatestLlamaAar(context) call before classloading
• Added versioned preference gate (dynamic_libs/llama_lib_version)
• Added asset lookup fallback order:
  • dynamic_libs/llama-.aar.br
  • dynamic_libs/llama-.aar
• On refresh, clears unzip/optimized cache dirs so new code is loaded
• Bumped LLAMA_LIB_VERSION to trigger rollout on existing installs

3) Assets generated/validated

Generated updated split assets and Brotli outputs:

• assets/release/v8/dynamic_libs/llama-v8.aar.br (updated)
• assets/release/v7/dynamic_libs/llama-v7.aar.br (new)
• Split zip regeneration includes updated dynamic_libs/llama.aar:
  • app/build/outputs/assets/assets-arm64-v8a.zip
  • app/build/outputs/assets/assets-armeabi-v7a.zip

Validation performed

• ./gradlew :app:bundleV8LlamaAssets :app:bundleV7LlamaAssets ✅
• ./gradlew :app:assembleV8Assets ✅
• Verified generated zip contains:
  • dynamic_libs/llama.aar
  • bootstrap.zip
  • android-sdk.zip
  • plugin-artifacts.zip
• Deployed to Samsung device (arm64-v8a) at:
  • /sdcard/Download/assets-arm64-v8a.zip
• Verified device file checksum matches local build checksum ✅

Why this is needed

Without these changes:

• Debug builds may not consistently pick up refreshed llama assets.
• Existing app installs can continue using stale extracted llama.aar even after updating bundled assets.

This PR ensures local LLM updates are reliably applied for both debug and release workflows using normal developer commands.

Ticket

ADFA-3105