[RL][Phase 4] Bind the C++ model and implement the Python trainer

[utilForever]

Parent: #1056
Epic: #1052

Goal

Train the same C++ policy-value module from Python without defining a second model.

Scope

• Bind the C++ torch::nn::Module as a Python torch.nn.Module-compatible object
• Expose forward, parameters, device movement, and weight save/load
• Implement the minimal Python training loop with torch.optim
• Consume the same replay sample contract as C++

Non-goals

• A pure-Python model implementation
• torch.compile, DDP, or AMP
• A second checkpoint format

Acceptance criteria

• Python autograd populates gradients on the C++ model parameters.
• A Python optimizer updates those parameters.
• Loss decreases on the same tiny repeated batch used by C++.
• The Python trainer saves weights loadable by C++.

Verification

Add one focused pytest training and weight-round-trip smoke test.