fix: degrade to cached and image results when the search engines are down

[felladrin]

Description

Currently, when every text-search engine fails (a rate limit or CAPTCHA suspends a SearXNG engine for an hour to a day), the search fails outright, even if the exact same query has a cached result, and the AI response has nothing to ground on.

This PR does three things:

• Serves the cached results of the same query when the live search fails, flagged stale (a banner in the results list and a disclaimer in the LLM prompt, so the model knows the results may be out of date).
• Falls back to the image search when the text search fails and image search is on: the answer waits for the image results and grounds on their titles and URLs, tagged (image) in the prompt.
• Makes the server fail fast (1 request, no backoff) when every unresponsive engine is under a long suspension, since they will not recover within the retry window.

Where to start

Most of the diff is tests (11 new ones, +354 lines). The behavior changes are in client/modules/search.ts (stale fallback plus a read-time age check), client/modules/textGeneration.ts (the image-fallback grounding, and a guard so an older search cannot repopulate a newer one's grounding channel), server/webSearchService.ts (the fail-fast) and the banner in client/components/Search/Results/Textual/TextSearchResults.tsx.

How to test

1. Run npm test (the new tests cover the stale fallback, the image grounding and the fail-fast).
2. Run npm run lint.
3. Note: a real outage needs the engines rate-limited or CAPTCHA'd, which cannot be forced locally, so the degraded UI is covered by tests only for now (the cases are in the docs/failure-injection.md matrix).