Restore Modal training entrypoint (CPU image + game-aware wandb)

[cweill]

Summary

modal_app.py and tests/test_modal_app.py are empty (0 bytes) in main — PR #7 ("Slim Modal image to CPU torch + game-aware wandb project") deleted both (375 + 348 lines) instead of editing them. modal run modal_app.py is broken, and #7's titled improvements never shipped. This restores the matched pair from before #7 and applies that intent for real.

Motivation

• Broken main: Modal training is unrunnable (empty entrypoint). The nbi C4 robustness study had to restore a /tmp copy to run at all.
• Lost improvements from Slim Modal image to CPU torch + game-aware wandb project #7: runs logged to a hardcoded WANDB_PROJECT = "alphazero-tictactoe", which is why the Connect Four robustness runs landed in a tic-tac-toe project.
• Undetected because CI has no Modal coverage (tracked separately as alphago-geq).

Changes

• Restore modal_app.py + tests/test_modal_app.py from 833d1fd~1 (last known-good pair).
• CPU-only torch image via extra_index_url=https://download.pytorch.org/whl/cpu — sequential MCTS gets no GPU benefit and CPU wheels are far smaller/cheaper.
• Game-aware wandb project alphazero-<game> via _wandb_project_for_game, replacing the hardcoded tic-tac-toe project. Added a regression test.
• Generic Modal app name alphazero (it trains any game).

Testing

• pytest tests/test_modal_app.py → 7 passed.
• Real-modal import check: app alphazero, image constructs with the CPU index, project resolves to alphazero-connectfour.
• End-to-end Modal smoke run passed (1 iteration tic-tac-toe, exit 0): the CPU image built and trained on Modal, logging to project alphazero-tictactoe. wandb run: https://wandb.ai/cweill-self/alphazero-tictactoe/runs/1p4aa843

Review focus

• The CPU-torch image approach (extra_index_url) vs. pinning +cpu wheels.
• Whether the generic app name alphazero is preferred over per-game app names.

Fixes alphago-67y. Related: alphago-0oz (wire batched MCTS into training self-play), alphago-geq (Modal CI smoke test).