feat: add Mistral Voxtral TTS extension

[TiagoAgora]

Summary

Adds a new TTS extension, mistral_tts_python, integrating Mistral's Voxtral text-to-speech via the OpenAI-compatible /v1/audio/speech endpoint.

• Built on AsyncTTS2HttpExtension (HTTP TTS mode), following the extension development guide.
• Requests a self-describing WAV stream and converts it to PCM16 mono on the fly (handles int16/int24/int32 and IEEE-float payloads); audio output at 24 kHz.
• Forwards vendor params through unchanged (model, voice_id, ref_audio, …); API-key auth via the Authorization header.
• Handles cancellation/flush, content-moderation/auth errors, and TTFB metrics.

Testing

Standalone unit tests pass in the dev container (tman install --standalone + tests/bin/start): 13/13 passing — covering config defaults/validation, URL/base_url resolution, headers, audio dump, flush/cancel, invalid-key handling, reconnect robustness, metrics, and the request state machine.

Live API validation against api.mistral.ai has not been run yet.

[YiminW]

HI Tiago, please run tts-guarder for your new tts extension and add all cases passed sanp here

[TiagoAgora]

TTS guarder results — mistral_tts_python

Ran task tts-guarder-test EXTENSION=mistral_tts_python CONFIG_DIR=tests/configs against the live Mistral / Voxtral API.

14 passed · 1 skipped · 1 failed

⏭️ test_subtitle_alignment — skipped; Voxtral exposes no word-level timing.

❌ test_interleaved_requests — traced to a pre-existing bug in ten_ai_base (tts2.py), not this extension. When finish_request() releases buffered interleaved requests it pre-sets _processing_request_id, so _process_input_queue() skips the QUEUED → PROCESSING transition. The later QUEUED → FINALIZING is then rejected as an invalid transition, so tts_audio_end is never emitted and the request queue stalls until timeout (log: Invalid state transition … queued -> finalizing). With a one-line fix — transition to PROCESSING when the dequeued item is still QUEUED — the test passes in ~30s. This affects every HTTP-TTS extension, so I've kept it out of this PR. Happy to open a separate ten_ai_base issue/PR.

Three issues the guarder surfaced in this extension are fixed in this branch:

• httpx[http2] dependency — the client uses http2=True; without the h2 package it failed to initialize on every request.
• dump_path added to the basic_audio_setting{1,2} test configs (the guarder reads config["dump_path"]).
• mistral_tts_python added to the fixed-sample-rate allowlist in tests/bin/start (Voxtral emits a fixed 24 kHz, like openai_tts2 / humeai).