[RL][Phase 5] Implement bounded replay and an end-to-end self-play smoke test

[utilForever]

Parent: #1057
Epic: #1052

Goal

Retain recent valid self-play samples in the smallest bounded replay implementation needed by the MVP.

Scope

• Append complete episode samples
• Sample mini-batches for both trainers
• Enforce a deterministic capacity and eviction rule
• Preserve schema/version metadata
• Exercise episode generation, serialization, replay insertion, and sampling together

Non-goals

• Prioritized replay
• A database service
• Distributed writers
• Long-term archival

Acceptance criteria

• Capacity is never exceeded.
• Eviction and fixed-seed sampling are reproducible.
• C++ and Python consume replay batches with the same values.
• Interrupted or invalid episode writes do not add partial samples.
• One complete mirror-match episode reaches both trainer batch loaders.

Verification

Add focused replay tests and one end-to-end self-play-to-batch smoke test.