feat(webui): add catalog-backed reasoning effort selection - #5983

Problem

Reasoning effort is currently hidden under Advanced options as a free-text
field. Users have to guess supported values even when an authenticated provider
catalog already reports the selected model's capabilities.

Changes

• Move Reasoning effort directly below Model in the preset editor.
• Offer Default, the selected model's reported levels, and Custom.
• Preserve saved values such as xhigh; keep custom input focused when typed
  text temporarily matches a listed option or catalog data arrives.
• Distinguish provider-reported live, cached, and stale capabilities from
  built-in, preset/provider, and generic suggestions.
• Keep explicit default-only models from receiving invented effort choices.
• Normalize only recognized aliases of the effective provider's model prefix.
• Carry reasoning-capability provenance through Codex, xAI, and Copilot catalogs.
• Add translation keys across all ten supported locales and regression tests.

Ordinary OpenAI /models data is not treated as model-specific reasoning
confirmation. Providers without capability metadata retain clearly labeled
suggestions and custom input.

Scope

Preset editor and capability metadata only. The inference/save plumbing,
authentication, and chat model picker remain unchanged. Default uses the existing
empty-value save path; loading catalog data alone does not change the draft or
save configuration.

The Codex catalog-version fix is a separate, independent PR: #5984.

Verification

• 149 backend tests pass:
  tests/providers/test_oauth_model_catalog.py and tests/webui/test_settings_api.py.
• 182 UI tests pass across reasoning-effort control, settings models, catalog
  feedback, i18n (including locale parity), and API tests.
• WebUI build, changed-file Ruff, and ESLint checks pass.
• 18 isolated desktop/mobile browser checks pass, including option selection,
  custom-input focus, model switching, and no save caused by catalog loading.

Browser checks use mocked catalog responses and in-memory saves; they do not
establish live inference acceptance.