fix: clarify PaddleOCR OCR configuration UX

[webup]

What

• default the OCR page to the hosted PaddleOCR endpoint instead of an empty field plus a misleading placeholder
• stop reusing the Models-page AI Studio key in the OCR form, and add inline guidance that PaddleOCR needs its own token
• clear stale test-status messages when users edit or save the OCR form, and map OCR failures to actionable token/endpoint/network/service guidance
• add regression tests for the default endpoint, stale-error cleanup, and OCR error-message mapping

Why

• issue OCR 配置页面多项 UX 问题：endpoint 提示误导、token 自动填充错误、错误提示混淆、错误信息模糊 #152 reported four UX failures in the OCR setup flow: the endpoint prompt was misleading, the page reused the wrong token source, stale test errors looked like save failures, and OCR failures surfaced as generic HTTP 500 messages
• the current flow made first-time OCR setup harder than it should be and gave users very little help when the upstream service rejected a request

How

• introduced a hosted PaddleOCR default endpoint constant and used it to seed the OCR form
• removed the shared-model-token prefill path and tightened saved-config detection to require a real OCR token
• extracted OCR error translation into a small helper so the page can consistently convert raw transport/upstream failures into user-facing messages
• updated the OCR page to reset stale connection status on edits and saves, then added focused Vitest coverage

Verification

• npx vitest run src/modules/ocr/__tests__/OcrPage.test.tsx src/modules/ocr/__tests__/errorMessages.test.ts
• npm run build
• dev-browser verification against the real PaddleOCR service provided in the issue follow-up: live connection test succeeded and a real sample parse rendered markdown in the patched OCR page

Screenshots

• UI was verified locally against the patched OCR page and the real PaddleOCR service.
• I could capture the page locally during verification, but this CLI environment could not upload the image into the GitHub PR body automatically, so this PR is left as draft until a screenshot is attached in GitHub.

Closes #152