Currently we randomize the validator we use to download blobs and, if the chosen validator fails, fail over to another in round-robin fashion. However, this is vulnerable to starvation if the validator is very slow or, as seems to currently happen on the Web, the underlying transport retries a lot, and currently leads to extremely poor initialization times in some circumstances (up to minutes!).

Instead, we should try several (maybe even all?) validators concurrently, and only proceed with a subset that respond in a timely fashion.