Implement per-provider v3 catalogs

[r4victor]

Closes #251

• Each provider job now collects, integrity-tests, and publishes its own catalog, so a failing provider no longer blocks the others. Layout: v3/{provider}/{version}/catalog.zip, v3/{provider}/latest/catalog.zip, and v3/{provider}/version
• The client fetches providers in parallel, keeps the previously loaded catalog for a provider that fails, and skips the download when the published version is unchanged.
• v2 catalogs are still published unchanged for existing dstack installations, and still only when every provider succeeds. The datacrunch.csv alias moved into the v2 job, since only v2 clients need it.
• v1 catalogs are dropped.
• The online-provider integrity tests became separate test-hotaisle and test-vastai jobs, and a final report job writes per-job status to the run summary.
• Collection failures are sent to a Slack channel.
• CI uses uv instead of pip, following the dstack setup.

Breaking changes

Catalog.get_latest_version() is removed, as per-provider catalogs have no single version. load(version=...) now requests that version from every provider and raises if any provider lacks it, rather than mixing versions. GPUHUNT_CATALOG_URL is now a base prefix rather than a full catalog URL.