fix: return Gymnasium-style dict infos from MarkovVectorEnv

[teddytennant]

Problem

pettingzoo_env_to_vec_env_v1 wraps a PettingZoo parallel env in MarkovVectorEnv, which is a gymnasium.vector.VectorEnv. On reset and step it returned infos as a Python list of per-agent dicts:

infs = [infos.get(agent, {}) for agent in self.par_env.possible_agents]

Gymnasium vector environments (since v0.25) expect infos to be a dict of arrays with boolean _<key> masks (see VectorEnv._add_info). Wrappers such as gymnasium.wrappers.RecordEpisodeStatistics (and the vector variant) assert isinstance(infos, dict) and fail with:

AssertionError: `info` dtype is <class 'list'> while supported dtype is `dict`.

This breaks CleanRL multi-agent PPO (ppo_pettingzoo_ma_atari.py), which wraps the SuperSuit vec env with RecordEpisodeStatistics.

Root cause

MarkovVectorEnv never adopted the post-v0.25 Gymnasium vector info layout. Downstream ConcatVecEnv / ProcConcatVec also assumed list-of-dicts and flattened nested lists, so the whole SuperSuit vector stack was list-shaped.

Fix

• Add supersuit/vector/utils/info_dict.py with list_infos_to_dict (list of per-env dicts -> Gymnasium vector dict) and merge_vector_infos (concat blocks along the env axis).
• MarkovVectorEnv.reset / step now return dict infos. Non-agent (global) PettingZoo info keys are still propagated into every agent slot. Terminal auto-reset still merges step + reset infos with reset keys winning.
• Set metadata["autoreset_mode"] = AutoresetMode.SAME_STEP to match the existing same-step auto-reset when the underlying PZ env finishes.
• ConcatVecEnv and ProcConcatVec merge vector info dicts instead of flattening lists.
• SingleVecEnv returns dict infos for consistency.
• SB3VecEnvWrapper converts dict infos back to a list of per-env dicts, which is what Stable-Baselines3 expects.

Test plan

• test_infos_are_dict_not_list - reset/step return dict
• test_record_episode_statistics_compatible - gymnasium.wrappers.vector.RecordEpisodeStatistics works and emits episode / _episode
• test_infos_dict_through_concat_vec_envs - same through concat_vec_envs_v1
• Existing test_pettingzoo_to_vec.py suite updated for dict infos (terminal_observation masks, etc.) - 9 passed
• Related vector tests (test_gym_vector, test_vector_dict, vec_env_test) - passed / skipped as before

pytest test/test_vector/test_pettingzoo_to_vec.py -o addopts=''

Fixes #249