Fix Whisper models hanging/failing to load while Parakeet works fine

[Dopixel]

Every Whisper model load called WhisperKit.download(), which performs a live Hugging Face Hub round-trip to list and verify remote files, even when the model is already fully downloaded on disk. If that request stalled, loading would hang indefinitely with no timeout and no way to fall back to the local files already present - reproducible by simply re-selecting an already-downloaded model from the menu bar.

• Skip the Hub round-trip entirely when the model's local folder is already known, loading WhisperKit directly from disk instead (this is what Parakeet/FluidAudio already does, which is why only Parakeet worked reliably).
• Wrap the remaining WhisperKit init in a generous timeout, since first- time loads still require a real (sometimes slow, several-minute) on-device Neural Engine model compilation plus a tokenizer fetch from a separate Hub repo - so a genuinely stalled request now fails with a clear error instead of spinning "Loading..." forever.
• Fix a dormant matching bug in the "is this model downloaded" folder scan: it compared variant strings for exact equality against on-disk folder names, but downloaded folders are always prefixed with "openai_whisper-" while the variant string only sometimes is (broken for base.en, small.en, and large-v3 specifically).