feat: migrate common/data/columnar module

[lxy-9602]

Purpose

No Linked issue.

Migrate common/data/columnar/ module. This module provides columnar data access layer built on top of Apache Arrow arrays, including:

• ColumnarArray: Access vector column data for array types
• ColumnarMap: Access vector column data for map types
• ColumnarRow: Row view over arrow::Array with per-row field access
• ColumnarRowRef: Lightweight row view sharing batch-level context to reduce per-row overhead
• ColumnarUtils: Utility functions for reading Arrow arrays (including dictionary-encoded arrays)
• ColumnarBatchContext: Shared batch context struct for ColumnarRowRef

Also migrates the dependency DateTimeUtils (timestamp conversion utilities).

Tests

• columnar_utils_test.cpp — Tests GetView/GetBytes for plain and dictionary-encoded string arrays
• columnar_array_test.cpp — Tests all primitive types, complex/nested types (struct, list, map, decimal, timestamp), and null handling
• columnar_row_test.cpp — Tests ColumnarRow and ColumnarRowRef for simple types, complex/nested types, dictionary types, null handling, data lifecycle, and binary access

API and Format

Documentation

Generative AI tooling

[leaves12138]

Thanks for the update. The non-owning ColumnarArray constructor is clearer now, but I found two remaining correctness blockers in nested row lifetime and decimal conversion.

[leaves12138]

This nested row still drops the ownership chain. The returned ColumnarRow stores raw child Array pointers, but it is constructed without a holder. If the parent ColumnarRow is the only object owning the top-level StructArray, a nested row returned from GetRow can outlive the parent and then its raw child pointers dangle. The current DataLifeCycle test only uses result_row while the parent row is alive, so it does not catch this. Please keep an owning holder here, for example by passing a shared_ptr/sliced StructArray or by returning ColumnarRowRef with a ColumnarBatchContext that owns struct_array->fields().

[leaves12138]

Please avoid reconstructing the signed int128 value by left-shifting a signed high word. For negative Decimal128 values, decimal.high_bits() is negative, and left-shifting a negative signed integer is undefined behavior in C++. This same expression is also used in ColumnarRow and ColumnarRowRef. Please combine the bits through an unsigned 128-bit value first, then cast to Decimal::int128_t, and add a negative decimal test.

[leaves12138]

I did another full pass and found one additional correctness issue beyond the two existing blockers.

[leaves12138]

This only decodes dictionary indices for INT32 and INT64, but Arrow dictionary indices are not limited to these two widths (for example int8/int16 dictionaries are valid and commonly used for small dictionaries). For those arrays dict_index stays -1; in release builds the assert is compiled out and the code then calls dictionary->GetView(-1), which can read before the offsets buffer or crash. Please handle all supported integer index types (or return a proper error instead of relying on assert).