fix: schema-aware Arrow struct accessor for partial nested projections

[wombatu-kun]

Fixes #499 — partial projection of nested struct children (e.g. SELECT s.b, s.c FROM t over struct<a, b, c, d>) crashed the Lance vectorized reader with UnsupportedOperationException from ArrowVectorAccessor.getLong.

• Lance's native scan does not push down nested struct projection, so the Arrow StructVector always carries all on-disk children in physical order; LanceStructAccessor was binding by physical Arrow ordinal but Spark's generated projection (and external consumers like Hudi's LanceRecordIterator) index by the pruned schema's ordinal — type mismatch.
• Adds schema-aware constructors LanceArrowColumnVector(ValueVector, DataType) and LanceStructAccessor(StructVector, StructType) that bind Arrow children to a Spark StructType by name, recursing into nested structs. LanceFragmentColumnarBatchScanner.loadNextBatch threads the input-partition schema through so lance-spark's own scan also uses the schema-aware path.
• Keeps ReadSchemaNestedStructWidening from fix: widen pruned nested struct schemas to preserve Arrow child ordinals #442 as defense-in-depth for the standard scan path.