feat(tts): single-model engine — swap CustomVoice/Base on demand

Replace dual-engine TTS (both models loaded, ~6.1 GB) with single-engine
that swaps variants on demand (~3 GB, saving 51%).

- Qwen3TtsEngine::swap_model() wraps Synthesizer::swap_talker()
- Merged cv_engine + base_engine into single engine + variant tracker
- ensure_customvoice/ensure_base swap in-place, cold-load only on first use
- CloneOneSentence recovers if variant swapped away mid-batch
- Eager load now loads only CustomVoice (default)
- Model status reports actual variant name

Tested: 10 interleaved cycles (20 swaps), no memory leak (+44 MB drift).

Closes #10

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

bump: v1.0.1

- Switch default TTS to 5-bit quantized models
- Rebase upstream MLX-C 0.4.0 fixes
- Extract qwen3-tts-core crate