v1.7.1: Gemma 4 & error annotations

What changed

• Gemma 4 model support — upgraded llama-cpp-2 to 0.1.143 with native Gemma 4, Llama 4, and 134+ architecture support
• Structured error annotations — inline error explanations with What/Cause/Fix sections and red line highlighting directly in the terminal
• Error annotations in Neuro tab — auto-detected errors are saved and browsable in the AI panel with per-session toggle
• Jinja template fallback — models with unrecognized chat templates now render correctly via Jinja/minja instead of failing
• Speculative decoding fix — resolved ISWA decode crash affecting Gemma 4 and other interleaved sliding-window models
• Agent panel polish — fixed hover states, button visibility, idle content wipe, and adaptive CPU timeout for slow prompt processing
• Key input lag fix — deferred highlighting to RAF flush and optimized ghost typing retries under high throughput