Add NHL94 self-play: asymmetric drill modes, opponent snapshot pool, curriculum integration

[Copilot]

Implements Milestones 1–4 of the NHL94 self-play plan: side-aware action processing, frozen opponent execution inside the wrapper, terminal zero-sum drill rewards, and checkpoint-pool rotation — all without touching PPO's single-agent interface.

nhl94_rf.py — Dedicated finetune reward modes

• SelfPlayOffenseFinetune: attack-zone init; terminal +1 on team-1 score, -1 if team-2 secures ≥60 consecutive frames of possession or clears the puck, 0 on timeout
• SelfPlayDefenseFinetune: defense-zone init; terminal +1 if team-1 clears past centre ice or holds possession ≥60 frames, -1 on team-2 goal, 0 on timeout
• Per-episode frame-counting state (_selfplay_offense_ctrl / _selfplay_defense_ctrl) follows the existing _defensezone_carry pattern; both modes registered in _reward_function_map

nhl94_obs.py — Side-aware action state + self-play interface

Replaces the three shared button-debounce fields with a per-side dict so learner and opponent never corrupt each other's state:

self.action_state = {
    "learner": {"b_pressed": False, "c_pressed": False, "slapshot_frames": 0},
    "opponent": {"b_pressed": False, "c_pressed": False, "slapshot_frames": 0},
}

• _process_filtered_action(ac, side) / _process_multidiscrete_action(ac, side) extracted as helpers
• Removes dead prev_state.Flip() (was unreachable in the reward/obs path)
• New public methods: set_opponent_model(path), set_selfplay_role(role), compute_opponent_action(obs), combine_selfplay_actions(learner, opponent)
• step() injects the frozen opponent's processed actions when selfplay_enabled=True; team 1 is fixed as the learner throughout Milestone 1

train_live.py — Opponent snapshot rotation

• New args: --load_opponent_model, --selfplay_snapshot_freq, --selfplay_pool_size
• OpponentSnapshotCallback saves a snapshot every N steps, maintains a bounded pool, and samples the next opponent with a 40 % latest / 40 % random historical / 20 % oldest distribution (degrades gracefully for small pool sizes)
• LiveTrainer seeds the initial opponent at startup via env_method("set_opponent_model", ...) and attaches the callback when --selfplay is active

train_curriculum.py

• Added load_opponent_model to PATH_KEYS so relative paths in the curriculum JSON are resolved correctly

curriculum/nhl94.json

• Appended four alternating self-play phases after the existing subskill curriculum: Offense Finetune 1 → Defense Finetune 1 → Offense Finetune 2 → Defense Finetune 2, each with selfplay: true and snapshot/pool config