[VL] Add task barrier support for NativePlanEvaluator to make Velox task reusable

[zhztheplayer]

Description

Currently ColumnarBatchOutIterator returned by NativePlanEvaluator#createKernelWithBatchIterator is disposable. It could only serve one input stream and output the result data one single time.

We'd add support for Velox's task barrier to make it reusable.

Fixing this will significantly improve performance of the ongoing native Delta statistics writer PR where we are creating a new Velox task for every single written partition, which can become super slow when there are thousands of partitions to write.

Reference, Velox task barrier:

https://facebookincubator.github.io/velox/develop/task-barrier.html

Gluten version

Main