fix(lance-table): treat trailing empty segments as no excess in rechunk_sequences

[LuciferYang]

Closes #9301

rechunk_sequences errors when segments are left over after the last chunk is filled. Its fill loop drains empty segments only while a chunk still needs rows, so a trailing empty segment survives to that check and a call whose chunk sizes sum to exactly the number of row ids fails with InvalidInput. The check now asks whether a leftover segment still carries ids.

Empty segments are already tolerated everywhere else in the function, with a test case for each position: inside a chunk, at a chunk boundary, and several in a row. Trailing was the one gap. The outcome changes only for inputs where every leftover segment is empty, including the degenerate case of an all-empty input against empty chunk_sizes; the check is not gated on allow_incomplete, so this applies in both modes. A segment the last chunk only partially consumed is not empty, so leftover ids are still reported.

No caller reaches this today. Production row removal goes through RowIdSequence::mask, which drops empty segments, and the sequences the four callers pass in come either from rechunk_sequences itself or from a capture stream whose scan nodes drop zero-row batches. The capture path has no zero-row guard of its own: extract_row_ids captures every batch it polls, and RowIdSequence::from(&[]) yields one empty segment, so the invariant lives in the upstream nodes rather than at the capture site.

How was this patch tested?

test_row_id_sequence_rechunk_with_empty_segments gains three cases: two trailing empty segments with the row counts matching exactly must produce one chunk of [0, 1]; a segment with ids behind a trailing empty must still be reported as excess; and so must a segment the last chunk only partially consumed. All were mutation-checked. With the old check the first case fails with "Got too many segments for the provided chunk lengths. Processed 1 chunks out of 1 expected"; with a check that never reports excess, the other two fail along with the pre-existing excess case in test_row_id_sequence_rechunk.

[lance-gatekeeper]

✅ Gate recommendation: approve.

The terminal check now matches rechunk_sequences' row-count contract: leftover segments are rejected only when they still contain row IDs. The added regression cases cover consecutive trailing empties while confirming non-empty and partially consumed excess remains invalid.

[LuciferYang]

The windows-build failure is unrelated to this change: it is the recall assertion in index::vector::ivf::v2::tests::test_legacy_ivf_pq_cosine_multivec_smoke (recall 0.48 against a 0.5 threshold on randomized multivec data). That test builds a legacy IVF-PQ index and never reaches rechunk_sequences, and windows-build has been green on recent main commits. Everything else on this PR is green.