fix: bound a shuffle read's coalesced batches by bytes as well as rows

[msafonov]

Which issue does this PR close?

Closes #. (Joom fork, OOM diagnosis of night 10-17 -> 10-18, fix A)

Rationale for this change

ShuffleReadCoalescer joins decoded shuffle blocks until spark.comet.batchSize rows, with no bound on bytes, outside any memory pool. In prepare_device_events (stage 52: shuffle read of HLL sketches, about 190 KB per row, about 6 rows per block) that is about 1.5 GB per batch per task plus the concat_batches copy: allocated reached 16 GiB with 0 reserved within 4 s of the stage start, and 12 executors were OOMKilled (em32, limit 26g).

What changes are included in this PR?

• ShuffleReadCoalescer::new(target_rows, target_bytes): a batch is passed on at the row target or the byte target, whichever comes first; a block that alone exceeds the byte target passes through unchanged (flushing what is pending first).
• Bytes are the sliced size of each column (ArrayData::get_slice_memory_size), because a decoded block slices every column out of one IPC body and buffer capacity would count the body once per column.
• New config spark.comet.shuffle.read.coalesce.maxBytes, default 64 MiB, 0 = no byte bound. It reaches the JVM decoder through createShuffleReadCoalescer(batchSize, maxBytes) and the native ShuffleScanExec through the new ShuffleScan.coalesce_max_bytes proto field.
• Narrow rows are unaffected: 8192 rows stay under 64 MiB up to about 8 KB per row.

How are these changes tested?

• New Rust tests in read_coalescer.rs: stops_joining_wide_rows_at_the_byte_target (60 blocks of 6 x 190 KB rows: one batch of >60 MiB without the byte bound, batches <= 4 MiB with it, same rows in order), passes_a_block_larger_than_the_byte_target_through, counts_a_body_shared_by_all_columns_once. With the byte checks disabled all three fail.
• Local (aarch64 container): cargo fmt --check, cargo clippy --all-targets --workspace -D warnings, workspace tests green (core 661, shuffle 193, spark-expr 972, jni-bridge 28, common 70).
• Cluster validation on the datasci-ab playground follows (prepare_device_events vs vanilla).

🤖 Generated with Claude Code