### What problem does this PR solve?

Issue Number: None

Related PR: apache#65396

Problem Summary: Ordered reads of nested struct fields over a rowset with
multiple overlapping segments use pruned sub-column iterators, but the merge
reader creates its intermediate blocks from the full physical schema. The
mismatched child layouts can decode strings into numeric columns, return the
wrong fields, or crash the BE. Backport the merge-block portion of apache#65396 to
branch-4.2: use the FE-pruned type by column unique ID when constructing merge
blocks, retaining the physical type for unpruned columns. Include focused
unit and regression coverage for block reuse and nested field reads.

### Release note

Fix errors and BE crashes when ordered scans read pruned complex columns
across overlapping segments.

### Check List (For Author)

- Test:
    - ASAN BE build and FE build.
    - Unit test: VGenericIteratorsTest.*, all 7 tests passed.
    - Regression test: fault_injection_p0/test_topn_pruned_struct_overlapping_segments.
      The original merge-block construction reproduced E-3110 for TINYINT;
      the same suite passed all 4 queries after restoring the fix.
    - Regenerated the expected output with the regression script; it matches
      the original PR's corresponding generated output byte for byte.
    - clang-format 16 passed. clang-tidy found no diagnostics on changed lines;
      its full run is blocked by an existing unmatched NOLINTEND in core/types.h.
- Behavior changed: Yes. Merge blocks use the requested pruned complex types.
- Does this need documentation: No

Partial backport of 3c43f4c (merge blocks).