Fix probabilistic interleave of empty sources without replacement

[HaokaiDing]

Fixes #8626.

When probabilistic interleave_datasets uses all_exhausted_without_replacement, an empty source starts as not exhausted. Its first draw appends an invalid source offset, producing a duplicate row from the next dataset or an IndexError when the empty source is last.

Mark empty sources exhausted initially for this strategy, so the existing skip-exhausted-source logic handles them. The other stopping strategies are unchanged.

This is separate from #8399 (no-probabilities all_exhausted) and #8318 (vectorized probabilistic first_exhausted / all_exhausted); the latter retains the without-replacement loop with the all-False initialization.

Validation:

• Before the fix: 9 empty-source cases fail across seeds 0, 42 and 1234; 5 non-empty/all-empty controls pass.
• After the fix: HF_HUB_OFFLINE=1 .venv/bin/pytest tests/test_arrow_dataset.py -k 'probabilities_without_replacement_empty_source or probabilities_without_replacement_all_empty' -q: 14 passed.
• Ruff lint, Ruff formatting and git diff --check pass.
• The full test suite was not run.

Prepared with OpenAI Codex; tests were executed locally and the diff received an independent review by another Codex agent.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.