Fix DiffSinger rendering (Phonemizer & Multi-Speaker) & Misc.

[KahazaTester]

Pull Request: DiffSinger Engine Fixes, Singer Import Auto-Detection & L10n Fallback

This should fix the following issues: #277 #280

Summary of Changes

1. DiffSinger Renderer & Phonemizer Fixes

• Eliminate np.dot Crashes (DiffSingerSpeakerEmbedManager.cs):
  • Replaced np.dot matrix multiplication in both PhraseSpeakerEmbedByFrame (rendering) and PhraseSpeakerEmbedByPhone (duration model phonemization) with managed per-frame normalized weighting and direct float buffer copies, fixing NullReferenceException crashes on mobile runtimes.
  • Added path-segment suffix matching and fallback handling for speaker embeddings.
• Speaker Subbank Fallback (DiffSingerBasePhonemizer.cs):
  • Added fallback hierarchy in GetSpeakerAtIndex (tone+color match -> color match -> first defined subbank) to correctly resolve singers with named voice colors on default notes.
• Multi-Dict G2P Language Prefixing (DiffSingerG2pPhonemizer.cs):
  • Fixed multi-dictionary phoneme prefixing by checking !string.IsNullOrEmpty(langCode) instead of useLangId, ensuring voicebanks with use_lang_id: false correctly map G2P phonemes to <langCode>/<phoneme>.

2. Singer Import Auto-Detection (ClassicSingerSetupViewModel.cs)

• Added archive inspection heuristic when importing singer archives without explicit singer_type in character.yaml:
  • Detects dsconfig.yaml -> sets SingerType = "diffsinger".
  • Detects enuconfig.yaml -> sets SingerType = "enunu".
  • Defaults to "utau" otherwise.

3. Localization Default (LocalizationManager.cs)

• Changed SystemLanguageFallbackCode from zh-Hans to en so unrecognized system languages default to English on first start.

Verification

• Tested on Android 16 (NNAPI & CPU).