feat(settings): improve AI provider model discovery and error handling

[Frankli9986]

Summary

Improves the AI provider model-discovery experience in Settings:

• Adds a shared model-fetcher module with provider-specific /models callers.
• Surfaces fetch errors to the user instead of silently returning an empty list.
• Allows manual model-name entry as a fallback.
• Remembers the selected model per provider.
• Migrates legacy persisted provider configs to a versioned shape.
• Adds unit tests for the fetcher, route handler, and settings store.

Changes

Backend

• New src/lib/ai/model-fetcher.ts with normalizeBaseURL and fetchModels.
• src/lib/ai/provider.ts: exports AIProvider, PROVIDER_DEFAULTS, getDefaultBaseURL, getDefaultModel.
• src/app/api/ai/models/route.ts: delegates to fetchModels, returns { models, error }.

Frontend

• src/stores/settings-store.ts: versioned per-provider config persistence with legacy migration.
• src/components/settings/settings-dialog.tsx: loading/error/retry states, AbortController for request races, manual model input, aria-* attributes.

i18n

• messages/en.json / messages/zh.json: added settings.ai.apiKeyRequired, fetchModelsFailed, manualModel, retry.

Tests

• src/lib/ai/model-fetcher.test.ts
• src/app/api/ai/models/route.test.ts
• src/stores/settings-store.test.ts
• src/lib/ai/provider.test.ts

Design doc

See the workspace design spec for full context (internal reference).

Notes

• Does not add DeepSeek support; that will follow once PR fix(chat): support DeepSeek reasoning models by preserving reasoning_content #72 lands.
• If PR fix(chat): support DeepSeek reasoning models by preserving reasoning_content #72 is merged first, this branch can be rebased on the new main.

[vercel]

@Frankli9986 is attempting to deploy a commit to the twwch's projects Team on Vercel.

A member of the Team first needs to authorize it.