Highlights:
- OpenAI realtime streaming model gpt-realtime-whisper, surfaced
  under the existing openai backend (no separate --backend value).
  Streams partials live as you speak, identical UX to Vosk.
- New --type-direct flag (and Options → Keyboard backend → Type
  directly): bypass the clipboard and synthesise raw keystrokes —
  works in terminals and any other app without depending on a paste
  shortcut. Caveat: xkb-layout-dependent for non-ASCII.
- Auto-pick is now cloud-first: groq → openai → whisper → vosk.
- README + docs/ restructured: short landing page, in-depth content
  split into per-topic pages under docs/.
- VoskTranscriber migrated to StreamingSTTBackend (shared with
  desktop-ai-core>=0.2.0).
- Pasting-into-terminal callout: Ctrl+Shift+V workaround.

Fixes:
- Groq picked as default backend without explicit --model no longer
  raises 'Unknown backend' (latent bug exposed by the cloud-first
  auto-pick order).