QQ platform sends voice messages in Silk format (or special AMR
variants) that ffmpeg cannot decode, causing STT to fail.
- ensure_wav(): detect Silk format and convert via tencent_silk_to_wav()
- ensure_wav(): detect AMR, try ffmpeg first, fallback to convert_to_pcm_wav()
- _get_audio_magic_type(): broaden Silk detection to handle \x02 prefix
and SILK bytes anywhere in the first 16 bytes of the header
f8c48ce4cc