Add mobile-focused local LLM choices

[pgotta]

What changed

• replaces the old Qwen2.5 1.5B, Llama 3.2 3B, and Phi-3.5 Mini local downloads
• adds Qwen3 4B Instruct 2507 as the strongest overall local choice
• adds Gemma 3n E4B Instruct for nuanced long-form summaries and writing
• adds LFM2 2.6B as the smallest and fastest local choice
• shows a short best-for/tradeoff blurb and formatted download size before downloading
• upgrades the Llamatik/llama.cpp Android runtime to 1.8.1
• checks available storage and hardens resumable multi-gigabyte downloads
• removes obsolete local-model files from older app versions

Uninstall and storage behavior

Models and partial downloads remain in Sonario's private filesDir/models directory. Android removes this app-specific storage when the app is uninstalled. The manifest also explicitly disables fragile-data retention, and models/ is excluded from cloud backup and device-to-device transfer so multi-gigabyte model files are not restored after reinstall.

Important runtime note

This updates the model choices and the llama.cpp wrapper, but it does not add a Qualcomm-specific NPU backend. Llamatik requests available acceleration and may fall back to CPU depending on the phone and runtime build.

Version

Sonario 1.5.0.