Desktop patches Python bootstrap, GPU rendering, and model discovery

The desktop app no longer needs a Local Server first run to make Open Terminal work. Previously, startOpenTerminal threw 'Python is not installed' when Python was missing, which only produced a generic error toast and made the feature appear broken. Now it automatically triggers installPython(), reusing the existing download/extract/uv-install flow [1].

Linux rendering gets a targeted fix: both --disable-gpu and --disable-gpu-compositing were killing the display compositor, leaving <webview> guest surfaces as gray rectangles. The replacement, --in-process-gpu, "sidesteps the cross-process shared-memory IPC crashes while keeping the compositor alive so webview content renders" [2].

Hugging Face model downloads were invisible to llama-server because its --models-dir scanner only checks one level of subdirectories; models stored at models/huggingface/<repo-slug>/<filename> were two levels deep. The cache directory now points to models/ so downloads land one level deep, with automatic migration of existing models from the legacy subdirectory [3].