v0.8.0

Added

• Voice input (server). A new TranscriptionBackend Protocol (async,
  owner-scoped transcribe) with a NullTranscriptionBackend default (voice
  off → 410), resolved from DJANGO_AG_UI["TRANSCRIPTION_BACKEND"] via
  resolve_transcription_backend. get_urls(view, transcribe=backend) mounts a
  TranscribeView at <prefix>transcribe/ (POST a multipart audio clip →
  {"text": "<transcript>"}), server-validated by TRANSCRIPTION_MAX_BYTES /
  TRANSCRIPTION_ALLOWED_TYPES and carrying the same require_authenticated /
  get_user auth seam as the agent endpoint. The opt-in
  django_ag_ui.contrib.transcription.openai_transcription_backend.OpenAITranscriptionBackend
  is a ready reference impl over any OpenAI-compatible /audio/transcriptions
  endpoint (lazy openai import via the [openai] extra; subclass to change the
  model or base_url). New exports: TranscriptionBackend,
  NullTranscriptionBackend, TranscribeView, resolve_transcription_backend.
• Model reasoning forwarding. When a reasoning model is configured to think
  (via MODEL_SETTINGS), its chain-of-thought now streams to the client as the
  standard AG-UI reasoning events Pydantic-AI emits — a pure pass-through (no
  protocol extension), and the run transcript ignores the ephemeral events so
  nothing reasoning-related is persisted. The new FORWARD_REASONING setting
  (default True) gates it: set False to let the model reason privately while
  the events are stripped from the stream.