Make model downloads resilient across regions

[ppXD]

What changed

• add automatic Hugging Face ↔ configurable mirror failover, remembering the endpoint that succeeds
• retain partial model files and resume with HTTP Range after drops or app restarts
• add bounded retries, connect/read timeouts, and HTTP/SOCKS proxy support
• add Settings → Model downloads with live source, mirror, and proxy configuration in English, Simplified Chinese, and Traditional Chinese
• apply one shared policy to ASR, diarization, denoise, Core ML, and embedding model downloads

Why

A single Hugging Face request currently fails for users in regions where that endpoint is slow or unreachable. Interrupted multi-hundred-megabyte transfers are also deleted, forcing users to restart from zero.

Validation

• root workspace tests pass
• app tests: 30 passed
• wisp-models: 51 passed, 1 opt-in real-network smoke ignored by default
• wisp-embed: 21 passed, 8 existing real-model tests ignored
• root, app, and embedding clippy pass with -D warnings
• Svelte check: 0 errors; production Vite build passes
• real mirror download smoke passes
• official and mirror SenseVoice token files are byte-identical (sha256:f449eb28dc567533d7fa59be34e2abca8784f771850c78a47fb731a31429a1dc)