Upgrade speech recognition models

[IchenDEV]

Summary

• migrate authoritative Apple Speech results to the macOS 26 SpeechAnalyzer API while retaining legacy live preview and a safe fallback
• pin and automatically migrate the local Qwen ASR runtime to qwen3-asr-mlx==0.1.1
• expose Volcengine SeedASR 2.0 as the recommended model while preserving BigASR 1.0 compatibility
• move WhisperKit to the maintained argmax-oss-swift package and refresh the MLX dependency resolution

Why

The speech stack had several stale upgrade paths: existing Qwen virtual environments were treated as ready without checking their installed version, Apple transcription still relied entirely on the legacy recognizer, and Volcengine only exposed its older resource. The dependency metadata also still referenced WhisperKit's former repository.

Impact

Existing Qwen runtimes upgrade in place when needed. Existing saved Volcengine resource IDs remain unchanged, while new installs default to SeedASR 2.0. Apple Speech uses the newer analyzer for final file transcription and falls back to the established implementation if model assets or analysis fail.

Validation

• swift test — 459 XCTest tests passed, 7 skipped, 0 failures; 2 Swift Testing tests passed
• Apple SpeechAnalyzer integration test passed with the bundled Chinese sample
• Qwen runtime integration test passed against the migrated 0.1.1 environment
• swift build
• bash scripts/build-and-run.sh --verify
• git diff --check and localization plist validation