perf: increase SortPreservingMergeExec prefetch buffer from 1 to 16

When SPM reads directly from I/O-bound sources (e.g., DataSourceExec
without SortExec buffering), the merge loop stalls waiting for Parquet
I/O on each poll. Increasing the prefetch buffer lets background tasks
read ahead while the merge processes previous batches.

Local benchmark (release, 16 partitions, sort_pushdown_sorted):
  Q1 full scan:   Main 110ms → PR 82ms (1.3x faster)
  Q3 SELECT *:    Main 239ms → PR 228ms (1.05x faster)
  Q2/Q4 LIMIT:    3-7ms (unchanged, already fast)

Author: zhuqi-lucas

datafusion/physical-plan/src/sorts/sort_preserving_merge.rs

@@ -361,7 +361,7 @@ impl ExecutionPlan for SortPreservingMergeExec {
                     .map(|partition| {
                         let stream =
                             self.input.execute(partition, Arc::clone(&context))?;
-                        Ok(spawn_buffered(stream, 1))
+                        Ok(spawn_buffered(stream, 16))
                     })
                     .collect::<Result<_>>()?;