[RL][Phase 5] Define the self-play sample format

[utilForever]

Parent: #1057
Epic: #1052

Goal

Define the minimum versioned training sample produced by self-play and consumed by both trainers.

Scope

• Store observation, legal mask, normalized MCTS visit-policy target, outcome, and schema metadata
• Define dtype, shape, ordering, and player perspective
• Provide deterministic serialization and validation
• Keep the format independent of trainer language

Non-goals

• A general-purpose dataset framework
• Long-term migration across arbitrary schemas
• Storing complete search trees

Acceptance criteria

• C++ and Python read the same sample values.
• Policy targets are normalized and assign no mass to illegal actions.
• Outcomes use the observation player's perspective.
• Invalid versions, shapes, masks, or non-finite values are rejected.

Verification

Add C++ and Python round-trip tests using one small golden sample.