Which issue does this PR close?

Closes #2307

Rationale for this change

Reduce task serialization size and memory usage by avoiding serialization of the full rddPartitions array per task.

Previously, each task closure captured a reference to the upstream NativeRDD, causing the entire rddPartitions array to be serialized with every task. This PR introduces a NativePartition[P] wrapper that embeds only the relevant partition payload into each task's closure, and marks rddPartitions as @transient so the full array is excluded from serialization.

What changes are included in this PR?

• Add @transient to rddPartitions in NativeRDD, with numRddPartitions stored for lightweight partition reconstruction after deserialization
• Add NativePartition[P] case class to carry original partition payloads through NativeRDD without re-indexing into the partitions array
• Add NativePartition.unwrap() helper for safe partition extraction (handles reuse-exchange scenarios where the partition may not be a NativePartition)
• Wrap input partitions in NativePartition before NativeRDD creation across all NativeExec operators
• Store original NativePartition array in AuronShuffleDependency.rddPartitions for Spark 4.x ShuffleWriteProcessor compatibility
• Add @transient to inputFileScanRDD in NativeFileSourceScanBase (large driver-only field)
• Add @transient to private constructor parameters in Exec classes where the field is only used on the driver (NativeBroadcastExchangeBase.mode, NativeParquetInsertIntoHiveTableBase.cmd, NativeHiveTableScanBase.basedHiveScan)

Are there any user-facing changes?

No.

How was this patch tested?

UTs.