fix(transcription): make self-hosted model configurable

[xAlcahest]

Summary

In self-hosted mode the model was never configurable: getTranscriptionModel() in audioManager.js has no self-hosted branch, so it resolved the model from leftover cloud provider settings and for a fresh setup always fell back to gpt-4o-mini-transcribe, which most self-hosted servers reject with model_not_found. The self-hosted panel now has a Model field (stored as remoteTranscriptionModel) that is sent verbatim when set; left empty, everything behaves exactly as before.

Tracing the same path also showed that shouldUseStreaming() never checked self-hosted mode: a self-hosted user with a stored OpenAI key and the default realtime model was silently routed to OpenAI's realtime WebSocket instead of their own server. Self-hosted mode now always goes through batch HTTP to the configured URL.

Fixes #1013

Changes

• src/helpers/selfHostedTranscription.js: new helper with isSelfHostedTranscription and resolveSelfHostedTranscriptionModel
• src/helpers/audioManager.js: getTranscriptionModel returns the self-hosted model when set, shouldUseStreaming returns false in self-hosted mode
• src/helpers/transcriptionAuth.js: shouldSkipTranscriptionApiKey delegates to the shared predicate
• src/stores/settingsStore.ts: new remoteTranscriptionModel key with setter
• src/hooks/useSettings.ts: expose the new key and setter
• src/components/SelfHostedPanel.tsx: optional model input, rendered only when wired
• src/components/SettingsPage.tsx: pass the model field through the transcription section
• test/helpers/selfHostedTranscription.test.js: node --test coverage for the helper