feat(huddle): add German Whisper transcription

[visipass-dev]

Summary

• replace the English-only Parakeet STT archive with pinned, multilingual Whisper Tiny int8 encoder/decoder/token files
• verify every downloaded model file by SHA-256 and install the model atomically with its MIT attribution sidecar
• add a Huddle transcript-language selector for German and English, restarting only the local STT pipeline when it changes
• pass the selected language into sherpa-onnx Whisper with task = "transcribe"

Validation

• just ci
• real local German audio smoke test: 6.55 seconds of 16 kHz mono speech decoded in 0.39 seconds and produced a German transcript

The model remains fully local. No cloud speech service or audio upload is introduced.