Optimize array_min, array_max for primitive types #21100
Description
Is your feature request related to a problem or challenge?
In the current implementation, for each row in the batch we construct a PrimitiveArray for the row, feed it to Arrow min / max, and then collects the resulting ScalarValues in a Vec. We then construct a final PrimitiveArray for the result via ScalarValue::iter_to_array. There's a bunch of overhead here: constructing N intermediate PrimitiveArray values plus the Vec, dynamic dispatch, Arc refcount bumps, etc.
We can do better by just iterating over the flat values buffer directly: that's already a PrimitiveArray, so we can just invoke the Arrow compute kernel ourselves on the appropriate slice of the array for each row.
Describe the solution you'd like
No response
Describe alternatives you've considered
No response
Additional context
No response