fix: retain successful backend results after another engine fails

[roli-lpci]

When one backend raises before a slower sibling finishes, _search_sync() can discard that sibling's successful results and raise DDGSException. The collector now processes remaining submitted futures before extracting results. The executor was already waiting for those futures during shutdown; the existing result limit still applies. Collecting a late failure preserves any previously recorded error, including timeout classification.

Fixes #427.

Validation:

• Five offline engine-stub regression cases pass, covering pending successful results, duplicate prevention, all-backend failure, and preservation of earlier timeout/general errors. The primary regression fails against unchanged upstream code; the two error-preservation cases failed before the follow-up correction.
• make lint passes (Ruff and mypy, 33 source files).
• Final full pytest run: 18 passed, 4 failed. The same books/video integration tests also fail on unchanged upstream commit 70a5635; their failures are not introduced by this patch. Existing integration tests and their inter-test pauses are unchanged.