consider changing take for varbin

[a10y]

from a user report:

Trying to make sense of this: take on a chunked struct array ~ 600ms converting chunked array to arrow record batch and doing take ~ 25ms

what seems likely to be happening is that varbin.to_arrow is casting the varbin to a varbinview (relatively cheap) and then doing take on the views (extremely cheap).

we should consider killing the intrinsic varbin take impl and just converting to varbinview on take

[AdamGS]

We've also ran into ChunkedArray::take being really bad in the "shuffle" case where the indices aren't sorted, potentially creating O(len(take_array)) chunks.