feat: Gemma 4 on-device inference via LiteRT-LM (Android)

[sparkleMing]

Summary

Add support for running Gemma 4 models fully on-device on Android using the official LiteRT-LM Kotlin API.

What's new

Android native layer

• LiteRtLmPlugin.kt: Flutter platform channel plugin wrapping LiteRT-LM. Supports queue-based inference, streaming tokens, tool calls, thinking mode, model download via OkHttp, and M4A→PCM WAV audio conversion via MediaCodec.

Dart layer

• GemmaLocalClient: LLMClient implementation using platform channels. Acquires a global inference lock before each request.
• GemmaModelManager: Engine lifecycle manager. Vision/audio backends enabled strictly on demand — only when the request contains image/audio content. Engine fully torn down and rebuilt when backend config changes. Rebuild always happens after acquiring the lock to prevent teardown during active inference.

Provider integration

• New typeGemmaLocal provider type with model list (gemma-4-e2b, gemma-4-e4b)
• Model download UI in setup and settings pages (Android only)
• No API key or base URL required; LLM data sharing consent skipped for on-device models

Other fixes

• asset_analysis_tool: Gemma 4 uses JPEG + 896px max side to avoid LiteRT-LM patch count overflow. Non-Gemma path unchanged.
• pkm_skill / timeline_card_skill: Use state.metadata factId as fallback when model-provided fact_id is unreliable.

Dependency upgrades

• drift 2.30 to 2.32.1, sqlite3_flutter_libs 0.5 to 0.6, drift_flutter 0.2 to 0.3

[matthewchan-g]

Consider setting experimental flag enableConversationConstrainedDecoding to true to see if that helps with function calls generated by Gemma 4.

[sparkleMing]

这个PR主要因为当时测试下来难以保证全链路的Agent体验，所以没有合并进main分支

[N3tn1la]

Hi! Just wanted to check in — is this PR still being actively worked on? Any plans to merge it into main? Would love to see on-device Gemma inference land in the app!