perf issues in google-cloud-bigquery-storage

[pjfanning]

P-06 (Arrow allocator):

• SimpleRowReader now takes allocatorBytes parameter (no default)
• BigQueryStorageSettings gains arrowAllocatorBytes field (default 512MB)
• Configurable via withArrowAllocatorBytes() or config key arrowAllocatorBytes (supports size suffixes like 256 MB)
• Threaded through ArrowSource → BigQueryArrowStorage → GrpcBigQueryStorageReader

P-07 (readMerged):

• Seq.reduce((a, b) => a.merge(b)) kept — it's Seq.reduce on Source.merge which is streaming, not buffering
• Fixed SimpleRowReader leak in readRecordsMerged — now wrapped in try/finally reader.close()

[He-Pin]

The SimpleRowReader created here is never closed — each batch allocates a fresh RootAllocator that is never released, leaking native memory over time.

readRecordsMerged now correctly wraps the reader in try/finally reader.close(), but this path does not. Consider mirroring the same pattern:

a.map { batch =>
  val reader = new SimpleRowReader(session.schema.arrowSchema.get, allocatorBytes)
  try reader.read(batch)
  finally reader.close()
}.mapConcat(c => c)

[pjfanning]

I've fixed this based on your feedback, thanks.