Slice arrays before normalizing to avoid allocating over the whole backing buffer

### Describe the problem

Reported during review of #5262: reading two visible values out of a sliced array allocated approximately **8 MiB**, because DataFusion normalizes the entire backing buffer before slicing. Slicing first reduced the largest allocation to 416 bytes.

This is a component allocation measurement, not an end-to-end query timing.

### Describe the potential solution

Slice before normalizing so the work is proportional to the visible range rather than the backing buffer. A benchmark demonstrating the allocation should land with the fix so the improvement is measurable and protected.

### Additional context

Reported by @sunchao on #5262. The other allocation observation from the same review, the per-file Iceberg schema helper that built a field-id map for every task even with no required fields, was fixed in #5262 itself: presence is now tested per required id, so an empty required set does no work.

Locate the array normalization and slicing entry points involved in the visible-range path, then reproduce the allocation described in the issue. Add a benchmark covering two visible values from a sliced array and verify that slicing first reduces allocation from the backing-buffer scale to the visible range.

Written by the indexing model from the issue text.