Answering my own question:

When I look at the actual throughput metric (numRecordsInPerSecond) of the Map functions, it is similar for both case 1 and case 3. My mistake was to infer something about the throughput based on the metrics shown in the overview. However, the number of records that go out in the upstream operator increases with the number of side outputs. I wrongly assumed that each unique record is only counted once, even if it sent to multiple outputs.