spark: a CTAS round trip drops the target's partitioning and provider #1366

ToSubstraitRel.convertCTAS (ToSubstraitRel.scala:664-676) serializes only the target's names, schema and create mode. Partition columns, bucket spec, provider and location are dropped. ToLogicalPlan.visit(NamedWrite) then builds the table from catalogTable (ToLogicalPlan.scala:613-641), which is always a managed, unpartitioned parquet table at $warehouse/$table. A CTAS round trip therefore creates a different table from the one requested, and reports success.

// CreateDataSourceTableAsSelectCommand for:
//   CREATE TABLE ctas_part USING orc PARTITIONED BY (p) AS SELECT 1 AS id, 'a' AS p
// command.table: partitionColumnNames = [p], provider = orc
// -> ToSubstraitRel -> proto -> ToLogicalPlan -> execute
spark.sessionState.catalog.getTableMetadata(TableIdentifier("ctas_part"))
// partitionColumnNames = [], provider = parquet

#1269 added this kind of rejection for InsertIntoHadoopFsRelationCommand and InsertIntoHiveTable, but not for the CTAS commands.

Suggested fix: in convertCTAS, reject a target with partition columns, a bucket spec, or a provider or location that NamedWrite cannot carry, until the target's layout can be represented.

Measured on main at 72e902e (Spark 3.5.4 and 4.0.2). Pre-existing.