[Spark] Writes land at their final paths before the job commits #27

Vortex writes each task's file straight to its final path. Spark's own file sources write under _temporary and move files into place when the job commits.

What goes wrong

• A reader running during a write lists and opens files that are still being written, or that belong to a job that later fails.
• If the driver dies before the job commits or aborts, the files of every finished task stay behind, and the next read includes them. A rerun then duplicates those rows.

Failed task attempts are handled by #22, which deletes their files on abort. A job that never reaches commit or abort isn't.

Where

VortexDataWriterFactory.createWriter and PartitionedVortexDataWriter.createWriterForPartition build the final path directly. VortexBatchWrite.commit only logs.

Fix

Write to a staging location under the output path, for example _temporary/<job id>/, which #20 already hides from reads, and move or rename the committed files in VortexBatchWrite.commit. On object stores a rename is a copy, so this costs a second write per file. Spark's committers have the same trade-off, and vortex-data#9658 inherits Spark's committer when it moves to FileFormat.