[sjdoane]

After concatenating a streaming dataset selected with take(4), set_epoch(1) can change its rows from [0, 1, 2, 3] to [0, 3, 6, 9]. The same problem affects skip() and explicit shuffling after concatenation.

Propagate shuffling through vertical concatenation's shard wrappers so the existing skip/take ordering restriction is honored. This preserves the selected rows while allowing buffer shuffling within that selection. Related to the epoch handling in #8707.

Validation on macOS with Python 3.12 and CPU PyTorch:

• New regression matrix: 16 failures and 8 passing controls before the fix; all 24 pass afterward.
• Iterable and distributed unit tests: 714 passed, 39 skipped, 6 integration tests deselected.
• Repository-wide make quality and git diff --check pass.

The new tests cover Python and Arrow sources, skip/take through a map wrapper, 1/2/4 shards, single/mixed concatenations, repeated epochs and explicit buffered shuffling. Full repository and hosted integration matrices were not run locally.

Implementation and tests were generated with OpenAI Codex; the checks above were executed locally.