chore: Support native writer Overwrite and ErrorIfExists modes

[comphead]

Summary

• Fixes Comet's native Parquet writer honoring SaveMode. Because CometExecRule replaces DataWritingCommandExec with CometNativeWriteExec, Spark's InsertIntoHadoopFsRelationCommand.run() — where the
  path-exists check, delete-before-Overwrite, and Ignore short-circuit live — never runs. Overwrite against a populated target was silently acting like Append.
• Ports the non-partitioned/non-catalog branch of Spark 3's doInsertion logic into CometNativeWriteExec.doExecuteColumnar(): ErrorIfExists + exists → throw, Overwrite + exists → delete, Ignore +
  exists → skip, Append → proceed. SaveMode is now threaded through CometDataWritingCommand.
• Adds 10 tests in CometParquetWriterSuite covering each mode against empty / populated / empty-DataFrame targets, plus a materializeAsCometSource helper (inline Seq.toDF produces a LocalTableScan
  that trips requiresNativeChildren and silently falls back to Spark's writer).

[andygrove]

LGTM. Thanks @comphead.

I did not try and do a comprehensive AI review on this one because this feature is marked as experimental and disabled by default. We can keep iterating and do a full audit later, IMO.

[comphead]

Thanks @andygrove we gonna improve it as we test