Improve select throughput by parallelising fetches

While fixing MB-22844 I have done some experimentation and I have found that we can improve throughput by inlining Fetches with the Projections, if the projections are parallized.

In my tests I can get a select * from `travel-sample` down to less than a second (0.98 typically) from 1.4s!

Clearly this only works when max-parallelism > 1 and if we have multiple data nodes, and tends to be washed out by pretty=true, but it's not detrimental when it does not have any positive effects