Feature - Distributed self-play framework

[Codex-Crusader]

Summary

Design and implement a distributed self-play framework so multiple workers can generate training data in parallel.

Motivation

Distributed self-play allows scaling training data generation and shortens iteration cycles.

Acceptance criteria / Definition of Done

• Implement a basic worker + aggregator design:
  • Worker(s) run self-play and write examples to a central queue or storage.
  • Aggregator combines examples into a central replay buffer.
• Provide at least one example using local multiprocessing (multi-process) and a documented plan for extending to networked workers (e.g., Redis queue).
• Document how to run multiple workers and how they submit data.

Implementation notes

• Start with multiprocessing for local parallelism.
• For networked solution, consider simple message queue (Redis / RabbitMQ) later — but not required in first PR.
• Ensure reproducibility and avoidance of duplicate seeds/IDs.

Complexity

High — longer-term project suitable for experienced contributors.