Replace conv trunk with a configurable residual tower

[cweill]

Summary

Swap AlphaZeroNet's two-layer conv trunk for a stem + N residual blocks, with channels and num_res_blocks constructor knobs (defaults 64 / 4). This is the trunk architecture real AlphaZero uses (minus normalization, see below) and lets the network scale depth/width per game.

Motivation

Residual connections let gradients flow straight through a deep trunk, enabling more representational capacity than the current 2-conv stack without vanishing-gradient problems. Depth/width are now tunable per game rather than hardcoded.

Design notes

• No BatchNorm (intentional). The training loop can emit a size-1 final batch, which crashes BN in train mode. The skip connection — not BN — is the load-bearing idea, so the block is Conv→ReLU→Conv + skip → ReLU. BN can be added later alongside drop_last in train.py.
• Signatures unchanged. forward/predict/predict_batch are identical, so MCTS and training need no changes; only the trunk internals differ.
• Old conv-trunk checkpoints won't load into the new module (different layer names); there are no committed checkpoints.

Testing

• pytest tests/test_network.py tests/test_train.py → 18 passed.
• New tests: configurable depth/width, stem-only (num_res_blocks=0), skip-path gradient flow, and config validation.
• Training smoke (tic-tac-toe, 39 self-play examples): loss decreases (3.19 → 3.16 over 8 epochs), no NaN/divergence — confirms gradients flow and the deeper net trains. Full convergence vs. the old trunk needs a real training run.

Review focus

• Default num_res_blocks=4 / channels=64 — reasonable for small boards, or too heavy?
• The no-BatchNorm decision.