RandomizedContraction can leave checkpoint directories behind after a failed run #894

Problem

RandomizedContraction creates a per-run checkpoint directory and deletes it only after the output has been materialized successfully. Its finally block currently unpersists DataFrames and deregisters _axpb, but does not remove the temporary checkpoint directory.

If an intermediate operation fails, checkpoint data can remain in the filesystem or a remote object store. Repeated executions may therefore leak temporary storage.

This is related to the historical cleanup concern in #162, but specifically covers failed executions in the current RandomizedContraction implementation.

Proposed solution

• Reuse a scoped cleanup routine for the per-run checkpoint directory.
• Run it on both successful completion and from finally.
• Do not mask the original failure if cleanup itself cannot complete; log a warning instead.
• Add a regression test that fails after intermediate parquet materialization and verifies that no randomized-contraction-* directory remains.

The change does not modify public APIs.