feat(inspect): add ArrowRowBuilder for materializing Arrow batches

[WZhuo]

Summary

Adds ArrowRowBuilder (inspect/row_builder_internal), a schema-driven helper that materializes in-memory rows into an Arrow ArrowArray (a batch) for an arbitrary Iceberg schema. It wraps the nanoarrow boilerplate and exposes per-column access plus typed append helpers, so metadata tables (snapshots, history, manifests, …) can emit rows without re-implementing it.

This is the first of a series splitting metadata-table support into focused PRs; the InMemoryBatchReader and the SnapshotsTable::Scan integration are intended to follow in separate PRs that build on this.

What's included

• ArrowRowBuilder::Make/column/FinishRow/Finish and typed helpers AppendNull/AppendBoolean/AppendInt/AppendString/AppendStringMap.
• The implementation lives in the core iceberg library — it only needs nanoarrow + ToArrowSchema (no Apache Arrow), matching peers like manifest_adapter and arrow_c_data_util.
• Unit tests in row_builder_test.cc covering typed appends (int32/string/int64/boolean/map), null handling for optional columns, multi-entry/empty string maps, zero-row batches, and column-index bounds. Compiled into the metadata_table_test target.

Testing

• CMake (Ninja): cmake --build build --target metadata_table_test then ran it — 9/9 tests pass (4 existing MetadataTableTest + 5 new ArrowRowBuilderTest). ctest green.
• The test verifies output by importing the produced C-data into Apache Arrow (arrow::ImportRecordBatch), so its target is USE_BUNDLE.

Notes

• The test is registered under CMake's bundle build only. The meson build (which has no Apache Arrow/bundle layer) is left unchanged; the core-only metadata_table_test.cc continues to build there.
• Developed with AI-assisted tooling, reviewed by the author.