Adds a one-time download flow for the on-device Gemma 2B model (~1.1 GB)
with a progress bar, partial-download cleanup, and automatic navigation
to the app shell once the file is verified. App.dart now gates on
modelPresentProvider instead of onboarding state.
Adds a prompt builder that serialises 7-day mood/activity data, wires
InsightNotifier to call the on-device LLM directly (no user interaction),
caches results by prompt hash, and renders the analysis as a passive card.
Includes loading, empty (<2 entries), and error states with a force-refresh button.