Fix Ray deduplicator shared state

[macroguo-ghy]

Summary

• Share Ray deduplicator backend state across map_batches tasks for a single execution.
• Prepare Ray actor-backed dedup sets before serializing the operator into Ray tasks without recreating existing actor handles.
• Materialize Ray basic deduplicator stats for all stateful backends, including Redis, before later actions can re-run the lazy stats stage.
• Add regression coverage for document deduplication across Ray blocks, repeated executions, Redis materialization signaling, and actor handle reuse.
• Fix Ray test helper conversion by using RayDataset.to_list() instead of iterating RayDataset directly.

Fixes #971.

Validation

python3 -m pytest tests/ops/deduplicator/test_ray_document_deduplicator.py -q

Result: 9 passed, 10 warnings in 69.43s.

[gemini-code-assist]

Code Review

This pull request introduces a mechanism to handle stateful operators within Ray datasets by allowing operators to trigger dataset materialization after execution. This change specifically addresses potential issues in deduplication where Ray's lazy re-execution could lead to incorrect results due to persistent state in actors or external backends. The feedback highlights that the RedisBackend should also trigger this materialization to prevent similar state conflicts and suggests refactoring the actor initialization logic to eliminate code duplication.