feat(tts): add gandr_tts extension

[AALG123]

What

Adds gandr_tts, a TTS extension for Gandr (https://gandr.ai), under ai_agents/agents/ten_packages/extension/gandr_tts/.

Gandr exposes an OpenAI compatible speech endpoint, so the extension is a close mirror of the in tree rime_http_tts pattern: it extends AsyncTTS2HttpExtension from ten_ai_base and streams audio over a plain HTTP POST with httpx.

How it works

• POST https://tts.gandr.ai/v1/audio/speech with a Bearer API key and a JSON body of model, input, voice, response_format
• response_format is forced to pcm; the service streams s16le mono 24000 Hz and synthesize_audio_sample_rate() reports a fixed 24000
• Properties under params: api_key (defaults to the GANDR_TTS_API_KEY env var), voice (default gandr-mia), model (default tts-1), optional endpoint override
• Cancel/flush handling, error mapping (401 to INVALID_KEY_ERROR), dump support, and TTFB metadata follow the rime_http_tts implementation line for line

Gandr serves six voices (gandr-mia, gandr-ava, gandr-jenny, gandr-dane, gandr-leo, gandr-lewis) across 23 languages, and every render is watermarked. Keys are available at https://gandr.ai with a free tier of 50,000 tokens, so smoke testing against the live endpoint is straightforward.

Testing

• tests/ mirrors the rime_http_tts suite (basic dump/flush, error messages, TTFB metrics, params passthrough, robustness, state machine) adapted to the Gandr config surface; all vendor calls in these tests are mocked
• This was written against the in tree extension contracts and the Gandr API docs; I have not yet exercised it inside a full TEN graph, so I would appreciate a run of your extension test harness on this PR. Happy to iterate on review feedback and to sign the CLA.

Disclosure: I work on Gandr.