audio: add SpeechToText trait and OpenAI Whisper backend

[bug-ops]

Parent: #520
Depends on: #521

Context

Need a pluggable transcription abstraction so different STT backends (OpenAI Whisper API, local Whisper, future providers) can be used interchangeably.

Design

New module: crates/zeph-llm/src/stt.rs (or separate crate zeph-stt)

pub trait SpeechToText: Send + Sync {
    fn transcribe(
        &self,
        audio: &[u8],
        mime_type: &str,
    ) -> impl Future<Output = Result<Transcript, SttError>> + Send;
}

pub struct Transcript {
    pub text: String,
    pub language: Option<String>,
    pub duration_secs: Option<f32>,
}

pub enum SttError {
    UnsupportedFormat(String),
    FileTooLarge { size: usize, max: usize },
    TranscriptionFailed(String),
    NetworkError(String),
}

OpenAI Whisper backend

pub struct WhisperApi {
    client: reqwest::Client,
    api_key: String,
    model: String,  // "whisper-1"
}

• Uses POST /v1/audio/transcriptions with multipart form
• 25 MB file size limit
• Supports: flac, mp3, mp4, mpeg, mpga, m4a, ogg, wav, webm
• Language detection or explicit language parameter

Config

[audio]
enabled = true
backend = "whisper-api"  # or "whisper-local", "none"
language = "auto"        # or "en", "ru", etc.

[audio.whisper_api]
model = "whisper-1"

Acceptance criteria

• SpeechToText trait defined
• WhisperApi implementation with multipart upload
• Supported format validation
• File size check before upload
• Config section for audio settings
• Unit tests with mock HTTP responses