Support scribe_v2_realtime for live STT

[longevityboris]

Context

Per the public models doc (https://elevenlabs.io/docs/overview/models), ElevenLabs now ships two Scribe variants:

• scribe_v2 — batch transcription (current CLI default)
• scribe_v2_realtime — ~150ms latency streaming transcription

The CLI only exposes scribe_v2 and scribe_v1 via stt --model. Users running realtime ASR (captions, voice agents wiring their own stack) have to bypass the CLI.

What to do

• Add scribe_v2_realtime to the stt --model allowlist.
• Document it as the streaming option in STT_HELP + agent-info.
• If scribe_v2_realtime needs a different endpoint / response shape, wire that up.

Files

• src/cli.rs (SttArgs model value_parser)
• src/commands/stt.rs
• src/help.rs (STT_HELP)
• src/commands/agent_info.rs

Acceptance

• stt --model scribe_v2_realtime FILE hits the live endpoint and returns partial/final transcripts.
• Help / agent-info mention the model + its realtime-vs-batch trade-off.