bug: a panic in the shuffle transform aborts the build instead of returning an error

Description

IvfIndexBuilder::shuffle_data and reshuffle_partitions in rust/lance/src/index/vector/builder.rs run the transform for each batch in a spawned task and then unwrap the join result:

.buffered(get_num_compute_intensive_cpus())
.map(|x| x.unwrap())

A JoinError means the task panicked or was cancelled, so any panic inside the transform closure is re-raised in whatever task is draining the stream rather than returned to the caller of build().

There is a reachable panic in that closure. When a precomputed partitions file is in play, the closure attaches the partition ids with

batch = batch
    .try_with_column(PART_ID_FIELD.clone(), Arc::new(part_ids.clone()))
    .expect("failed to add part id column");

and try_with_column rejects a duplicate column name, so a batch that already carries __ivf_part_id panics here. The panic then travels through the unwrap above. That combination is what accelerator=... together with precomputed_partition_dataset=... used to produce.

Expected behavior

Attaching the partition ids should report a duplicate column as an invalid_input error naming the schema, and a panicking transform task should come back as an error from the shuffle rather than as a panic in the consumer.

Lance version

13.0.0-beta.4 (main)

Language binding

Rust

Start in rust/lance/src/index/vector/builder.rs with IvfIndexBuilder::shuffle_data and reshuffle_partitions, then trace the buffered stream's JoinError handling and the try_with_column call for PART_ID_FIELD. Reproduce the accelerator plus precomputed_partition_dataset case; done means duplicate partition columns produce an invalid_input error naming the schema and transform panics return an error from build() rather than aborting it.