[RL][Phase 4] Implement the native C++ trainer

[utilForever]

Parent: #1056
Epic: #1052

Goal

Train the shared policy-value model from replay samples using the LibTorch C++ frontend.

Scope

• Load mini-batches of observation, policy target, and outcome
• Compute policy loss, value loss, and the minimal documented regularization
• Run optimizer steps
• Save model weights and runtime-local training state as needed
• Report only essential losses and step counts

Non-goals

• Distributed training
• Mixed precision
• Hyperparameter search
• Python compatibility for optimizer state

Acceptance criteria

• Forward, backward, zero-grad, and optimizer steps run on CPU.
• Loss is finite and decreases on a tiny repeated batch.
• Checkpointed model weights reload correctly.
• Invalid shapes and non-finite samples fail before updating parameters.

Verification

Add one focused C++ tiny-batch overfit test.