Improve efficiency of `DatastoreIndexingRouter#source_event_versions_in_index`: leverage routing/timestamp

This method searches all shards an indices:

https://github.com/block/elasticgraph/blob/89b54463d1af5c0afa00aa8920b0a016c6a758b1/elasticgraph-indexer/lib/elastic_graph/indexer/datastore_indexing_router.rb#L173-L180

This is quite inefficient, particularly when you have many malformed events--it can cause significant load problems on a cluster.

The rationale given in that comment is speculative and I suspect we've never actually had malformed documents with the wrong routing/index rollover timestamp. We should make it more efficient.