feat: support VectorUDT writes via UDT unwrapping in LanceArrowWriter

[LuciferYang]

Summary

LanceArrowWriter.createFieldWriter had no case for UserDefinedType, so writing a DataFrame with a UDT column (most commonly MLlib's VectorUDT — the type produced by VectorAssembler, returned by many ML transformers) threw UnsupportedOperationException. This PR adds a single case that unwraps the UDT to its sqlType and recurses, matching the pattern Spark's native ArrowWriter uses. Read-back already worked because Arrow has no UDT concept — the column comes back as the underlying struct sqlType — so the write path is the only gap, and that's what this fixes.

Changes

One new case in LanceArrowWriter.createFieldWriter:

case (udt: UserDefinedType[_], _) =>
  createFieldWriter(vector, udt.sqlType, metadata)

Placed just before the catch-all case (dt, _) => throw new UnsupportedOperationException(...) so specific matchers run first and anything that's still a UDT at that point gets unwrapped. Metadata is threaded through so nested cases that need it — e.g. FixedSizeListWriter when a UDT's sqlType resolves to ArrayType with embedding metadata — still receive the right field metadata.

Tests

BaseSparkDataTypeRoundtripTest picks up testVectorUDTRoundtrip: creates a DataFrame with an id INT column and a vec VectorUDT column, writes two rows (a DenseVector and a SparseVector) as Lance, reads back, and asserts two things:

1. Schema contract. Arrow has no UDT, so the read-back schema must NOT carry VectorUDT — it must come back as the underlying sqlType (StructType). Asserts both halves: !instanceof VectorUDT and instanceof StructType. Locks this behavior so a future change that accidentally preserves UDT in field metadata is caught loudly.

2. Value fidelity. Reconstructs each row's vector from the read-back struct and checks equals against the original. A small reconstructVector(Row) helper does the manual reconstruction because VectorUDT.deserialize expects InternalRow, not the public Row that Spark hands back in test assertions. The helper switches on the type byte — 0 = sparse, 1 = dense — and throws IllegalArgumentException for anything else rather than silently misinterpreting bogus data as sparse.

A null-VectorUDT row is intentionally omitted and documented in the test's Javadoc: VectorUDT.sqlType marks its inner type field as non-nullable, and writing a parent-null struct produces null placeholders in the non-nullable children that Lance's native writer rejects. That's a Lance-side limitation with struct-of-non-nullable-children, not specific to UDTs; out of scope here.

spark-mllib is added as a test-scope dependency across all eight non-bundle modules (-base_2.12, -base_2.13, -3.4_2.12/2.13, -3.5_2.12/2.13, -4.0_2.13, -4.1_2.13). Each module pins its own spark<version>.version so cross-compile stays version-accurate. Bundle modules don't run tests so they don't need it. The 4.0/4.1 modules inherit the concrete SparkDataTypeRoundtripTest subclass via cross-compile from lance-spark-3.5_2.12/src/test/java, so the VectorUDT test executes on 4.0 and 4.1 too — spark-mllib is needed there for both compile and runtime.

[LuciferYang]

Thanks for your review. @jiaoew1991 ,can we merge this PR? I don't have merge permission.