[reginaldalfret]

Summary

Resolves #359

Modern reinforcement learning environments have largely transitioned from OpenAI Gym to Farama Foundation Gymnasium. Gymnasium introduces breaking API differences from legacy Gym:

1. reset(seed=..., options=...) returns (observation, info) instead of only observation.
2. step(action) returns 5 values (observation, reward, terminated, truncated, info) instead of 4 (observation, reward, done, info).
3. Space classes reside under gymnasium.spaces.

Passing Gymnasium environments into GymWrapper caused unpacking errors (ValueError: too many values to unpack (expected 4, got 5)).

Changes

• Implemented GymnasiumWrapper in acme/wrappers/gymnasium_wrapper.py:
  • Handles modern reset(seed=..., options=...) returning (observation, info).
  • Maps 5-tuple step returns into dm_env.TimeStep: terminated maps to dm_env.termination(reward, observation) (discount 0.0), truncated maps to dm_env.truncation(reward, observation) (discount 1.0), and normal transitions map to dm_env.transition(reward, observation).
  • Supports automatic reset when step() is called after episode termination/truncation.
  • Converts gymnasium.spaces (Discrete, Box, MultiBinary, MultiDiscrete, Tuple, Dict) into corresponding dm_env.specs.
  • Stores step and reset diagnostic metadata accessible via get_info().
• Exported GymnasiumWrapper in acme/wrappers/__init__.py.
• Gracefully handled optional legacy gym in acme/wrappers/__init__.py and optional launchpad in acme/utils/signals.py.
• Added unit tests in acme/wrappers/gymnasium_wrapper_test.py.

Verification

• Ran python -m unittest acme/wrappers/gymnasium_wrapper_test.py: 5 tests passed (covering spec conversion, reset with seeding, step transitions, termination, truncation, and auto-reset).