Improve performance of RunEndEncoded cast

[vegarsti]

Once #8709 is done, we'll have a benchmark we can use to try to improve the REE casting performance. @alamb and @tustvold have a neat suggestion to use the filter kernel:

I think what @tustvold is getting at is that the filter kernel is (very) fast and thus using it rather than the take kernel is likely to be faster.

So I think the idea here is instead of building up value_indexes it would be to implement some way to turn the result of Partition into a boolean array and pass to the filter kernel

    let partitions = partition(&[Arc::clone(cast_array)])?;
    // .... 
    let filter: BooleanArray = partitions.into_inner(); // not sure about this API
    // call filter instead of take
    let values_array = filter(&cast_array, &filter)?;

Originally posted by @alamb in #8589 (comment)

[vegarsti]

One small attempt, suggested by @alamb, is #8726