audio: real-time streaming STT support

[bug-ops]

Parent: #520
Depends on: #522, #523

Context

For real-time voice interaction, support streaming audio input with incremental transcription. This is a stretch goal for v1.

Options

1. OpenAI Realtime API — WebSocket-based, supports audio streaming with function calling
2. Local streaming — Whisper with chunked audio (VAD + sliding window)
3. Deepgram/AssemblyAI — third-party streaming STT APIs

Design

`SpeechToText` streaming extension

pub trait StreamingStt: SpeechToText {
    fn transcribe_stream(
        &self,
        audio_stream: impl Stream<Item = Vec<u8>> + Send,
    ) -> impl Stream<Item = Result<PartialTranscript, SttError>> + Send;
}

pub struct PartialTranscript {
    pub text: String,
    pub is_final: bool,
}

Integration points

• TUI: microphone input via `cpal` crate + VAD (voice activity detection)
• Channels: platform-specific streaming (if supported)

Acceptance criteria

• `StreamingStt` trait defined
• At least one streaming backend implemented
• TUI microphone input works (feature-gated)
• Partial transcripts displayed in real-time