DataShard: read actors retry TEvRead to shards long after they finished splitting

[snaury]

Looks like DataShard currently replies with a very generic error:

    if (!IsStateNewReadAllowed()) {
        replyWithError(
            Ydb::StatusIds::OVERLOADED,
            TStringBuilder() << "Shard " << TabletID() << " is splitting/merging"
                << " (node# " << SelfId().NodeId() << " state# " << DatashardStateName(State) << ")");
        return;
    }

Which causes ReadActor (LookupActor?) to keep retrying with an expotential backoff without resolving new partitioning. I've seen this causing 3+ seconds latency (due to expotential backoff) until it finally requests the new shard.

[snaury]

At the minimum DataShards should return NOT_FOUND after the split has finished, since the table no longer exists, and a new resolve attempt should return a new partitioning.