feat: Local TTS provider support (Supertonic, Index-TTS)

[fuleinist]

Problem

The current TTS pipeline (generate_voiceover) only supports remote HTTP API providers (bytedance, minimax, 302). Users who want to generate voiceovers locally — without relying on paid cloud APIs, rate limits, or network latency — have no option.

Local TTS is valuable for:

• Cost: zero per-character charges after the initial model download
• Latency: no round-trip to external APIs (~200–500ms on CPU vs. several seconds over the network)
• Privacy: audio never leaves the machine
• Offline/air-gapped environments
• Batch/render workflows where hundreds of segments make API costs prohibitive

Proposal

Add a new local TTS provider type alongside the existing remote providers. The implementation should:

1. Define a local provider interface (model path, voice, language, speed → WAV bytes)
2. Ship at least one built-in local engine
3. Allow users to select "Local" as the TTS provider in the web UI sidebar, choosing which local engine to use
4. Reuse the existing tts_providers.json parameter schema + config.toml [generate_voiceover.providers.*] pattern so local providers are configured the same way as remote ones

Suggested Local Engines

1. Supertonic (supertone-inc/supertonic)

• Repo: https://github.com/supertone-inc/supertonic
• Install: pip install supertonic (or pip install 'supertonic[serve]' for HTTP server mode)
• Model: auto-downloads to ~/.cache/supertonic3/ (~400MB)
• Languages: 31 (zh, en, ja, ko, fr, de, es, pt, ru, ar, hi, vi, etc.)
• Voices: 10 built-in styles (M1–M5 male, F1–F5 female), plus zero-shot voice cloning via Voice Builder JSON export
• Expressions: {{sad}}, {{laugh}}, {{whisper}}, {{shout}}, {{cry}}, etc.
• Output: 44.1kHz 16-bit WAV, no upsampler needed
• Hardware: CPU-only, runs on Raspberry Pi / edge devices. No GPU required.
• Speed: ~200–500ms per sentence on CPU (tested: 10.33s output in 9.61s wall time)
• SDK: from supertonic import TTS; tts = TTS(); wav, dur = tts.synthesize(text, voice_style=style, total_steps=8)
• HTTP server mode: supertonic[serve] exposes POST /v1/tts on 127.0.0.1:7788 — can reuse the same HTTP handler pattern as the existing 302/minimax providers

2. Index-TTS (index-tts/index-tts)

• Repo: https://github.com/index-tts/index-tts
• Description: Industrial-Level Controllable and Efficient Zero-Shot Text-To-Speech System
• Install: pip install index-tts
• Key feature: voice cloning from a reference audio file — provide a 10–30s clip and it replicates the speaker's voice
• Languages: zh, en, ja, and more
• Speed: GPU-accelerated (CUDA), with CPU fallback
• Model: downloads from HuggingFace (IndexTeam/Index-TTS)
• Use case: ideal for users who want a specific speaker voice without recording new audio

Acceptance Criteria

• User can select a local TTS engine (Supertonic and/or Index-TTS) from the web UI sidebar
• Local engine generates WAV output compatible with the existing voiceover pipeline (same voiceover_id / duration / path contract in generate_voiceover.py)
• Configurable via config.toml under [generate_voiceover.providers.supertonic] and [generate_voiceover.providers.index_tts]
• Parameter schema defined in resource/tts/tts_providers.json for each local engine
• Graceful fallback: if local model files are missing, prompt the user to run a download command rather than crashing
• No breaking changes to existing remote TTS providers