Tune nested take compaction threshold with a query-level benchmark #7145

What is the problem the feature request solves?

list_extract (#6498) and map_extract (#6845) share one policy for compacting nested take results: a buffer is compacted when its capacity exceeds 4× its used bytes plus 64. So far the threshold has only been measured at the kernel level. For 8,192 map<int,array<int>> rows with two 128-element values and about 75% missing keys, compaction makes m[1] 29–32% slower, while retained output capacity drops from 4,228,100 to 1,079,300 bytes (review). The kernel numbers miss the downstream effect: buffering operators such as sort reserve memory by capacity, so oversized output can raise memory pressure and spill earlier.

Describe the potential solution

Add a query-level benchmark that feeds nested m[k], element_at and arr[i] results into a buffering operator such as sort, and measures CPU time, peak memory and spills with ample and constrained memory. Use it to tune the shared threshold in nested_compaction.rs, or keep the current one if the results support it.

Additional context

Follow-up to @andygrove's approval of #6845. Both kernels should keep one policy.