[kevmo314]

Add Supertonic 2 as a smaller downloadable alternative to Supertonic 3: 10 speakers, English/Korean/Spanish/Portuguese/French, and a 267.3 MB download. All 16 upstream files were downloaded, size/hash verified, and pinned to revision 75e6727618a02f323c720cba9478152d4bc16ca4. Attribution and the model's OpenRAIL-M restrictions use the existing license display.

Installed packs now have an explicit, persisted radio selection. Supertonic 3 stays the default when both packs are installed; deleting the selected pack falls back safely. This avoids the old alphabetical selection silently switching everyone to the narrower-language pack. Engine replacement retains the existing utterance queue/disposal rules.

Validation:

• All 510 app Flutter tests and 96 on-device-TTS package tests pass; targeted Dart analysis and formatting pass.
• Actual Android emulator inference with the downloaded Supertonic 2 model: 2.675 seconds of audio synthesized in 6.693 seconds, nonzero RMS 0.034.
• Actual iOS 27 simulator native tests: both voice enumeration/provider fallback and English/Spanish synthesis pass (2 tests, 0 skips). The tests verify finite audio, duration, RMS, and deterministic/random seed behavior.

Research for #1731 also covered Qwen3-TTS. It has viable mobile runtime projects, but its autoregressive talker/code-predictor/codec pipeline needs a new native backend; it cannot use this app's existing Supertonic ONNX pipeline. Quantized GGML options are roughly 0.8–1.2 GB for the 0.6B model. This PR adds the verified compatible pack and selection foundation; it does not list Qwen3 as installable before that backend exists.

Addresses #1731.

Sources: Supertonic2, official Qwen3-TTS, Qwen3-TTS GGML runtime/model measurements.