Add Go (7x7) with ko, suicide rule, and area scoring

[cweill]

Summary

Adds Go — the project's namesake and most involved game — to the game-agnostic pipeline. 7x7 by default, with a pass action, captures, suicide prohibition, ko, two-pass termination, and Tromp-Taylor area scoring.

Motivation

Go stresses the Game interface in ways no prior game did: a non-spatial pass action (action_size = size² + 1) and history-dependent legality (ko). Implementing it on the same interface — with no changes to MCTS, the network, or training — is the strongest evidence the pipeline is genuinely game-agnostic.

Rule choices (documented in the module)

• Captures/suicide: place, remove opponent groups with no liberties, then reject the move if the mover's own group has no liberties (suicide illegal).
• Ko: simple single-point ko, not full positional superko. Full superko needs the whole position history in every state, which would bloat/slow every MCTS node; simple ko keeps GoState small and hashable and covers the common case.
• End/scoring: two consecutive passes end the game; Tromp-Taylor area scoring with a half-integer komi (default 7.5) so results are always decisive (no draws).
• Board size and komi are constructor args (default 7x7).

Testing

• Full suite: 141 passed.
• tests/test_go.py (12 cases): captures, plain suicide (illegal) vs capturing-suicide (legal), the ko ban + its clearing on a pass, two-pass termination, empty-board komi result, and territory attribution in area scoring.
• Integration smoke: game_from_name("go") → AlphaZeroNet → MCTS yields a valid 50-action policy (including pass).

Review focus

• The simple-ko vs positional-superko trade-off.
• Default board size (7x7) and komi (7.5) — reasonable, or prefer different defaults?
• legal_moves does a place+flood-fill suicide check per empty point (O(points²)); fine at 7x7, a known hot spot if scaled up.

Closes alphago-xdh.