Voice messages: speech-to-text -> prompt

[artickc]

Handle message:voice: download the OGG, transcribe via a Whisper-compatible endpoint (configurable in .env), then submit as a prompt. See src/bot/handlers/photo.ts for the download pattern.