fix(derived data): Adjust auto-heal timeouts in response to operational realities

[kcons]

We're working on a more efficient approach here, but in the meantime, we recognize that running these long queries in a task with a relatively short deadline can be a problem.

Here we allow the task to run longer, give the (cachable) hash discovery query enough headroom to reliably suceeed even when not recently vacuumed, and add a timeout that we don't expect to hit to our group id batching.

The total deadline isn't enough for worst-cases everywhere, but due to discovery caching, if we can find one or two stale hashes and cache that, that's enough to make progress.
Similarly, on the group batching side, we need to be able to process at least one, ideally two hashes to make significant forward progress each run.

One key motivation here: we always prefer to hit statement timeouts over task timeouts. Evidence indicates that task timeouts don't abort the running query, not remotely, and not locally, and this seems correlated with task failures going unreported. OperationalErrors within the task deadline seem to report reliably.