fix: rebuild the shuffle writer per iteration in end-to-end benches

[andygrove]

Which issue does this PR close?

No issue filed. Reported in #6095 (comment).

Rationale for this change

Every end-to-end bench in native/shuffle/benches/shuffle_writer.rs panics during criterion warmup:

shuffle write error: partition offsets were already published

Each bench builds one ShuffleWriterExec outside b.iter() and executes it on every iteration, but the writer publishes its partition offsets through PartitionOffsets, a OnceLock that errors on a second set. CI does not catch this because pr_benchmark_check.yml only runs cargo check --benches.

What changes are included in this PR?

• A bench_end_to_end helper that uses b.iter_batched to build a fresh exec for each iteration. Construction counts as setup and is not timed.
• The four end-to-end bench loops now call that helper instead of repeating the same execute block.

How are these changes tested?

Ran cargo bench -p datafusion-comet-shuffle --bench shuffle_writer -- --quick "end to end". All 12 end-to-end benches now finish (compression variants, hash/range/single partitioning, and 200/2000/8000 partitions). Before this change, all of them panicked.

[andygrove]

Folded this into #6095 instead, as the last commit on that branch. That PR cannot be benchmarked
at all until this lands, so splitting them just means measuring #6095 against a branch that has to
carry this anyway. The change is unchanged from what was here.

Closing in favour of #6095.