fix: useModels triggers unnecessary refresh=true scans on every mount

[cswaney]

Context

Follow-up to #467. useModels sends refresh=true unconditionally on every mount (loaders.js:14), which forces the backend to do a full filesystem scan + Hub metadata fetch on every tab open, focus event, and remount across three call sites (ModelsContainer, ServiceModal, NewJobModal). This was the fuel for the duplicate-row race fixed in #467, and it's independently wasteful backend work.

Why the default can safely change

• The backend inserts into the model table on download success (asgi.py:2626-2700). Models downloaded through the app appear in the DB immediately — no scan required.
• Without refresh=true, the endpoint returns DB contents. That's the correct source of truth for the common case.
• The models page already has a refresh button — the one legitimate "force a scan" trigger already exists in the UI.

The only case the DB-only read misses is out-of-band additions (someone runs hf download directly on the cluster). Users who hit that case can press the refresh button on the models page.

Proposed change

In useModels (loaders.js:8-40):

1. Drop refresh=true from the default SWR key and fetcher URLs.
2. Add revalidateOnFocus: false and a modest dedupingInterval (e.g. 60s) to the SWR config.
3. Expose a refresh() action that calls the refresh=true URL once and writes the result into the cache via mutate(fresh, { revalidate: false }).
4. Wire the existing models-page refresh button to refresh() instead of plain mutate().

No changes needed to ServiceModal or NewJobModal — they get the DB-cached list, which is fast and correct.

Expected impact

• Filesystem scans run only when a user explicitly presses the refresh button.
• Modal opens over high-latency filesystems (SSH / Open OnDemand login node) become instant on second-and-later opens instead of always re-scanning.
• Removes the last remaining trigger for the Race in get_models(refresh=true) creates duplicate Model rows #467 race pattern.

Verification

• Open a modal twice in quick succession → only one network call, no scan on the second.
• Tab-away-and-back → no scan.
• Download a model through the app → appears in modal without pressing refresh.
• Press refresh button on models page → forces scan, cache updates everywhere.