fix(writer): unify nullable utf8/binary on the NullableData carrier

[dfa1]

Summary

Follow-up to #167. Tackles the nullable-input shape asymmetry (utf8 used String[]-with-nulls, primitives used NullableData) by unifying on the NullableData carrier — and fixes two latent bugs that asymmetry was hiding.

What changed

• ChunkImpl — nullable Utf8/Binary now produces NullableData(String[]-with-nulls, validity), like nullable primitives.
• VarBinEncodingEncoder — a null element encodes as a zero-length slot (masked values child), no more NPE; stats already skip null.
• ZstdEncodingEncoder — consumes NullableData for utf8/binary; acceptsNullable covers them.
• CsvExporter — handles MaskedArray (null row → empty field).

Latent bugs fixed

• Default-path nullable utf8 write threw NullPointerException in VarBin.
• Nullable utf8 with no nulls in the scanned rows decoded as a bare VarBinArray, silently dropping null info; now consistently MaskedArray.
• CSV export of any nullable column (primitives too) failed with "unsupported array type".

Behavior change

Nullable utf8/binary now always decode as MaskedArray (validity + values child), consistent with nullable primitives. Tests updated (parquet importer, parquet import IT).

Testing

• Full unit suite (all modules) green.
• Full integration suite green, incl. new javaWriter_rustReader_masked_nullableUtf8 (Rust reads a Java-written nullable utf8 file) and the existing zstd-nullable ITs.

🤖 Generated with Claude Code