feat(transcription): add voice/video-note transcription

[PathKnower]

Stacked on #196 (small, unrelated list_messages media-label fix) — this
branch is #196 plus the transcription feature on top. The diff below will
shrink to just the feature once #196 merges; happy to rebase if you'd rather
review it standalone.

Summary

• New transcribe_voice(chat_id, message_id, engine=None) tool: transcribes a voice message or video note. Two engines behind one interface:
  • groq (default): uploads the recording to Groq's hosted whisper-large-v3-turbo. Costs a download+upload per call and leaves the server, but doesn't truncate long recordings.
  • telegram: native messages.TranscribeAudioRequest. Free, audio never leaves Telegram, requires Premium on the account — but in side-by-side testing on real recordings it silently dropped the last spoken clause in roughly 1 of 3 samples (confirmed: native output ended mid-sentence where Groq's output on the identical file continued for another clause). Long recordings return pending; the tool polls automatically instead of returning truncated/empty text.
• Engine choice is configurable two ways: the engine= argument per call, or TELEGRAM_TRANSCRIBE_ENGINE as the server-wide default.
• Results are cached by (chat_id, message_id) in a dedicated SQLite file (own directory, 0600/0700 permissions) since Groq transcription is a real per-call cost, not free like native — a repeat read/listing never re-transcribes the same message.
• get_history, get_messages, and list_messages now surface already-cached transcripts inline instead of leaving voice messages as empty text. TELEGRAM_TRANSCRIBE (off / on-demand default / auto) controls whether listings only show what's cached or also prefetch missing ones, bounded by a per-call budget (TELEGRAM_TRANSCRIBE_MAX_VOICES / _MAX_SECONDS) so one large listing can't silently trigger dozens of paid downloads.
• Every transcript is returned/rendered with an explicit note that it's a machine transcript, not a verbatim quote — wording drifts under both engines, so it shouldn't be treated as an exact quote of what someone said.
• transcribe_voice is readOnlyHint=True-annotated, so it stays exposed under TELEGRAM_EXPOSED_TOOLS=read-only.
• Docker Compose now mounts a ./transcript_cache volume so the cache survives a rebuild (previously would've lived only in the container's writable layer).

This is a bigger surface than a typical PR here — new tool, new env vars, a paid third-party dependency by default — so happy to adjust scope/defaults (e.g. default to telegram instead of groq, make the feature fully opt-in via TELEGRAM_TRANSCRIBE=off, split cache/budget into a follow-up) if you'd rather land a smaller version first.

Test plan

uv run pytest tests/ -q
# 361 passed (80 new: tests/test_transcription.py, tests/test_voice_transcripts.py)

uv run pre-commit run --all-files
black....................................................................Passed
flake8...................................................................Passed

Deployed to a live server and walked through both engines against 11 real voice messages in a real chat (15s–159s recordings): all transcribed successfully, repeat calls served from cache (no second API call), and the native-vs-Groq truncation difference reproduced on one of the recordings as described above.