Add Pocket TTS as a lightweight CPU-only TTS engine

[Ahmed-Ezzat20]

Summary

• Adds Pocket TTS (Kyutai, 100M params, MIT license) as a third TTS engine alongside Qwen3-TTS and Chatterbox
• CPU-only — no GPU required, ~6x real-time on modern CPUs, no CUDA/torch version conflicts
• Supports zero-shot voice cloning from reference audio (requires gated model access)
• Includes 8 predefined voices: alba, marius, javert, jean, fantine, cosette, eponine, azelma
• Gracefully falls back to predefined voice when cloning model is unavailable
• English only (warns on non-English language, doesn't crash)
• New optional dependency: pip install "mazinger[tts-pocket]"

Changes

File	Change
tts.py	New _PocketTTSWrapper, _load_pocket_model, _create_pocket_voice_state, _synthesize_pocket; updated TTSEngine, load_model, create_voice_prompt
cli/_groups.py	Added "pocket" to --tts-engine choices, new --pocket-voice arg
cli/_dub.py	Pass pocket_voice to pipeline
pipeline.py	pocket_voice param on dub(), override ref audio when set
pyproject.toml	New tts-pocket optional extra (pocket-tts>=1.0)

Usage

# With predefined voice (no special access needed)
mazinger dub video.mp4 --tts-engine pocket --pocket-voice alba

# With voice cloning (requires HF gated access)
mazinger dub video.mp4 --tts-engine pocket --voice-sample ./reference.wav

# CPU-only, no GPU needed
mazinger dub video.mp4 --tts-engine pocket --device cpu --pocket-voice marius

Test plan

• Verified model loads and predefined voice synthesis works (alba voice, 4.64s output)
• Verified graceful fallback when voice cloning model is unavailable
• Verified CLI parser accepts --tts-engine pocket and --pocket-voice across all subcommands
• Verified pipeline.py passes pocket_voice through to create_voice_prompt()
• Test full dub pipeline end-to-end with Pocket TTS
• Test with voice cloning model (requires HF gated access)