polish(majdoddin): use ENV_REGISTRY for loading AlphaZero simulation environments

[Majdoddin]

Background (DI-engine pattern)

DI-engine already loads classes for collector and evaluator environments via ENV_REGISTRY (using get_vec_env_setting() in
DI-engine's base_env.py). However, the simulation environment class used by MCTS in AlphaZero policies was still loaded through
hardcoded imports.

Changes

This PR completes the registry-based approach by refactoring simulation environment class loading to use ENV_REGISTRY, matching the
pattern used by DI-engine for collector/evaluator environments.

• Modified train_alphazero.py to pass full_cfg and create_cfg to policy config
• Replaced hardcoded _get_simulation_env() in all 3 AlphaZero policies:
  • lzero/policy/alphazero.py
  • lzero/policy/gumbel_alphazero.py
  • lzero/policy/sampled_alphazero.py

Benefits

• Consistency: All environment types (collector, evaluator, simulation) now load their classes via the same registry
• Extensibility: New environments can be added without modifying core LightZero code
• Reduces code duplication: ~130 lines of hardcoded imports removed

Testing

Tested and verified with:

• ✓ Gomoku Gumbel AlphaZero
• ✓ TicTacToe AlphaZero
• ✓ TicTacToe Sampled AlphaZero

All tests passed successfully.

[puyuan1996]

Thanks for this excellent contribution, @Majdoddin! This is a great refactoring that significantly improves consistency by aligning the simulation environment loading with the DI-engine pattern.

We have just one quick question before merging: could you confirm that the changes are fully compatible with both the 'play_with_bot' and 'self_play' modes? We noticed your tests on Gomoku and TicTacToe and just wanted to verify if both scenarios were covered.

Once that's confirmed, this looks good to go.

Thanks again for the great work!

[Majdoddin]

Thanks! I can confirm compatibility with both modes.
I've tested the registry-based changes with both play_with_bot and self_play modes across multiple games.

All tests passed (5/5):
Sampled AlphaZero (ptree - Python MCTS):

• TicTacToe Sampled AlphaZero - Bot Mode ✓
• TicTacToe Sampled AlphaZero - Self-Play Mode ✓
• Gomoku Sampled AlphaZero - Bot Mode ✓

AlphaZero (ctree - C++ MCTS):

• TicTacToe AlphaZero - Bot Mode ✓
• TicTacToe AlphaZero - Self-Play Mode ✓

All tests ran for 90 seconds each to verify environment initialization, data collection, and training starts successfully.

Testing notes:

• Tested only configurations that pass on upstream main (other AlphaZero ptree configs have pre-existing bugs on main)
• Built and tested a ctree_alphazero to verify ctree compatibility
• Rebased PR on latest main for compatibility with recent changes

See attached test output files for full details.
test_results_main_baseline.txt
test_results_pr424.txt

[puyuan1996]

Thanks for this excellent contribution！