feat: add list type support to ArrowWriter

[jding-xyz]

Summary

ArrowWriter currently cannot serialize schemas containing list<T> columns. This PR adds full list type support to the writer, including nested list<struct<...>> schemas.

1. List type writer support (ArrowWriterHelper.swift, ArrowWriter.swift)

• ArrowWriterHelper.swift: Added .list case to toFBTypeEnum (returns .list FlatBuffer type enum) and toFBType (creates empty org_apache_arrow_flatbuf_List table).
• ArrowWriter.swift: Added ArrowTypeList handling in writeField, writeFieldNodes, writeBufferInfo, and writeRecordBatchData — mirrors the existing ArrowTypeStruct pattern but uses ArrowTypeList.elementField and NestedArray.values for the single child array.
• Also moved the nested-type recursion in writeBufferInfo and writeRecordBatchData outside the inner buffer loop, so it runs once per column rather than once per buffer (a latent correctness issue for types with multiple buffers).

2. Field-node ordering fix for nested types (ArrowWriter.swift)

writeFieldNodes iterates in reverse to match FlatBuffers prepend semantics, but wrote parent field nodes before recursing into children. This placed child nodes before their parent in the final vector, violating Arrow IPC's depth-first pre-order requirement. External readers (pyarrow) reject such streams with "Array length did not match record batch length".

Moved child recursion (for both struct and list types) before the parent fbb.create call so children are prepended first and end up after their parent in the resulting vector.

3. Schema message padding in streaming format (ArrowWriter.swift)

The Arrow IPC spec requires all message metadata to be padded to a multiple of 8 bytes. writeFile already does this via addPadForAlignment, but writeStreaming did not. Without padding, the record batch body can start at a non-8-byte-aligned offset, causing alignment violations when strict readers (pyarrow + Rust FFI) create zero-copy buffers from the stream.

Added addPadForAlignment(&schemaData) in writeStreaming to match writeFile behavior.

4. Fix appendAny bypass in nested builders (ArrowArrayBuilder.swift)

The base ArrowArrayBuilder.appendAny calls bufferBuilder.append directly, bypassing the overridden append methods in StructArrayBuilder and ListArrayBuilder that distribute values to child builders. This means list<struct<...>> columns produce struct arrays with empty children, since the struct's field builders never receive data.

Overrode appendAny in both StructArrayBuilder and ListArrayBuilder to delegate to self.append, ensuring child builder distribution runs correctly for nested types.

5. Make ArrowField initializer public (ArrowSchema.swift)

ArrowField is a public class with all public properties, but its initializer was internal. This prevents external consumers from constructing ArrowTypeStruct instances with explicit fields, which is needed for list<struct<...>> column schemas.

Made ArrowField.init(_:type:isNullable:) public.

Test plan

• Verified all existing tests pass (swift test — only pre-existing fixture-file-missing failures).
• Round-trip tested with list<float32> schema: encode via ArrowWriter.writeStreaming, decode via ArrowReader.readStreaming, verified schema fields and row count survive.
• Round-trip tested with list<struct<string, float32>> schema: encode via ArrowWriter.writeStreaming, decode via ArrowReader.readStreaming, verified schema fields and row count survive.
• End-to-end validated list<float32> and list<struct<string, float32>> streams against a pyarrow-based server (Arrow IPC streaming reader) — both accepted without errors.
• Tested alongside the timestamp nesting fix from fix: move timezone string creation before startTimestamp to avoid nesting assertion #148 (independent change).