fix(index): report shuffle-stage failures instead of panicking

[LuciferYang]

Closes #9404

Two steps of the same failure. Attaching precomputed partition ids to a batch that already has __ivf_part_id hit try_with_column's duplicate-name error behind an .expect, so it panicked inside the spawned transform task; the stream then unwrapped the resulting JoinError, re-raising that panic in the consumer instead of returning it from build(). The attach now returns an invalid_input error naming the batch schema, and both shuffle streams (shuffle_data and reshuffle_partitions) turn a JoinError into an error rather than unwrapping it.

The JoinError mapping is the safety net for any other panic in that closure, so it stays even though the one reachable panic is now handled at the source.

How was this patch tested?

test_shuffle_data_reports_duplicate_partition_id_column drives shuffle_data with a precomputed partitions file and a batch that already carries the partition column, and asserts the error names the attach. Restoring the .expect makes the test fail with the error the JoinError mapping produces, shuffle transform task failed: task N panicked with message "failed to add part id column: ...", which is the mapping doing its job; restoring both lines makes the panic escape the stream again.

[lance-gatekeeper]

✅ Gate recommendation: approve.

The change handles the reachable duplicate-column failure at its source and prevents both shuffle streams from re-panicking on transform-task failures, while preserving their existing concurrency and ordering. The focused regression test covers the invalid-input path and its diagnostic context.

[LuciferYang]

mac-build failed in lance-index::vector::kmeans::tests::test_hierarchical_kmeans_is_seeded_and_balanced, a size-balance assertion in the kmeans trainer. This PR only touches the shuffle streams in rust/lance/src/index/vector/builder.rs, so it cannot reach that test; it passes locally on this branch. Every other job is green.