Batch ingestion: preserve Arrow columnar batches beyond the Iceberg reader #20401

Description

PR #19510 added an opt-in Iceberg Arrow reader, but the native batch-ingestion boundary is still row-oriented. IcebergArrowInputSourceReader currently converts every Arrow ColumnarBatch row into a Map<String, Object> and MapBasedInputRow. InputSourceReader then exposes CloseableIterator<InputRow>, InputSourceProcessor loops over individual rows, and IncrementalIndex.add accepts one InputRow at a time.

This issue proposes carrying the columnar representation deeper into native batch ingestion so Arrow decoding gains are not lost to per-row materialization before indexing.

Motivation

Local JMH results after #19510 show the bottleneck moving downstream:

• Reader-only workloads are 1.23–1.58x faster, with a 1.36x geomean.
• End-to-end read -> incremental index -> segment persist is 1.02x faster .

Acceptance criteria

• Existing row-based input sources and ingestion specifications remain compatible.
• The batch path demonstrates a statistically significant end-to-end improvement on at least one representative wide-table workload without regressing the TPC-H lineitem case.
• Unsupported operations fall back clearly to the row path rather than changing results.