In a scenario where a Dataflow streaming job processes messages from a message bus that does not support exactly-once delivery, applies transformations, and then loads the results into BigQuery, how can you ensure exactly-once delivery semantics at an ingestion throughput of about 1.5 GB per second?