The ProjectionPlan::project_batch function is inefficient and noisy

[westonpace]

In take workloads we utilize a function called ProjectionPlan::project_batch. This function creates a OneShotExec that reads in the batch and a ProjectExec that applies the projection expressions. This is probably fine if there are actually projection expressions to evaluate. However, if we all we are doing is reordering or dropping columns (or especially if the projection is an identity projections 🤦) we are adding quite a bit of unneeded overhead.

To put this into context, in a recent random access benchmarking effort I found this to be responsible for about 8% of the latency even though there was no actual projection (it was just identity).

[Xuanwo]

Maybe I can take a look over this🙈

[westonpace]

Sure! Maybe a good first step would be creating some kind of take rows benchmark (with random rows) in ci_benchmarks (in the python folder).

The existing datasets in ci_benchmarks are probably fine so we just need the test.