feat: add SenseVoice ASR model and improved model ranking

[tmm22]

Summary

This PR adds support for Alibaba's SenseVoice ASR model and implements an improved model ranking system.

New Features

SenseVoice ASR Model Integration

• SenseVoice - Alibaba's ultra-fast multilingual ASR model
  • Supports Chinese, Cantonese, English, Japanese, Korean
  • 15x faster than Whisper with 96% accuracy
  • 234 MB model size using ONNX Runtime inference
  • Uses LFR (Low Frame Rate) feature stacking for efficiency

Dynamic Model Ranking System

• Added speed and accuracy properties to the TranscriptionModel protocol
• Models are now sorted using a geometric mean score: sqrt(accuracy × speed)
• Models excelling at BOTH speed AND accuracy rank highest
• Bonus (+0.1) applied for models with high accuracy (>=0.94) AND high speed (>=0.75)
• Sorting applied across Recommended, Local, and Cloud tabs

Files Added

File	Description
SenseVoiceTranscriptionService.swift	ONNX-based inference with LFR feature stacking
SenseVoiceTokenizer.swift	Token decoding for SenseVoice output format
SenseVoiceModelCardView.swift	UI card with speed/accuracy ratings
WhisperState+SenseVoice.swift	Model download, delete, and management
FastConformerFeatureExtractor.swift	Audio feature extraction for ONNX models

Files Modified

• TranscriptionModel.swift - Added SenseVoiceModel struct and protocol extensions
• PredefinedModels.swift - Added SenseVoice model definition
• ModelManagementView.swift - Added ranking algorithm and SenseVoice actions
• ModelCardRowView.swift - Added SenseVoice card rendering
• WhisperState.swift - Added service and routing
• project.pbxproj - Added onnxruntime-swift-package-manager dependency

Dependencies

This PR adds the onnxruntime-swift-package-manager package from Microsoft for ONNX Runtime inference.

Testing

• [x] Build succeeds
• [x] Model downloads successfully
• [x] Transcription works correctly
• [x] Model ranking sorts models as expected

Screenshots

The SenseVoice model appears in the Local models tab with speed/accuracy ratings displayed.

---

Summary by cubic

Adds SenseVoice multilingual ASR via ONNX Runtime and a new ranking that prioritizes models that are both fast and accurate. This brings much faster Asian-language transcription and better model recommendations across tabs.

• New Features

  • SenseVoice ASR integration (zh/yue/en/ja/ko), ~234 MB, up to 15x faster than Whisper.
  • LFR feature stacking, custom tokenizer, and greedy decoding with a fast feature extractor.
  • SenseVoice model card with download/delete/show-in-Finder and progress UI.
  • Ranking updates: added speed/accuracy to TranscriptionModel; geometric-mean score with a small bonus; applied to Recommended, Local, and Cloud.

• Bug Fixes

  • Added SenseVoice routing in AudioFileTranscriptionService and Manager to prevent cloud fallback.
  • More robust downloads and decoding: HTTP status validation and CTC collapse for cleaner text.

^{Written for commit 7cb1b5478408792a5ebe65afc61deb265fd0983e. Summary will update automatically on new commits.}