cirrusdoc: Recognize broken page responses

There are examples of production pages that aren't missing, but also are broken and don't manage to return any usable response. Probably the cirrusdoc endpoint should be updated to mark these in some way, but the SUP should also be resilient. If the API returns a page without the expected cirrusdoc throw an invalid response exception and give it a few retries before dropping it into the fetch failed topic.