When DingTalk sends a voice message to the bot, the callback JSON contains
a 'recognition' field with the speech-to-text result (powered by Qwen).
Previously, LangBot only extracted the 'downloadCode' to download the raw
audio binary and passed it as 'file_base64' to LLM APIs, which caused
400 errors since most models don't support this content type.
This patch:
- Extracts the 'recognition' field from DingTalk audio message content
- Uses it as plain text input to the LLM instead of raw audio
- Falls back to audio binary only when no recognition text is available
- Fixes duplicate text issue for audio messages with recognition
Fixes voice messages returning 'Request failed' on all LLM models.
