[Feature] Updated TD-MPC2 evaluation and fixed some bugs

[t-sekai]

#439

• Separated eval_env from train env (Align evaluation setups for different online RL algorithms #486)
  • eval_env is now used for evaluation. The only difference between env and eval_env is num_envs.
  • num_eval_envs in configs now defines the num_envs of eval_env.
• Fixed max_episode_steps (tdmpc2 baseline AttributeError: 'ManiSkillVectorEnv' object has no attribute 'max_episode_steps' #528)
  • Newer gymnasium and ManiskillVectorEnv doesn't contrain default max_episode_steps as an attribute. Therefore, default max_episode_steps is now determined by gym_utils.find_max_episode_steps_value(env).
• Fixed wandb logging to log exactly eval_episodes
  • Realized that wandb logging in evaluation wasn't logging exactly eval_episodes (it was a multiple of num_eval_envs). Futhermore, only num_eval_envs videos were logged at each evaluation.
  • It is now fixed to log exactly eval_episodes at each evaluation.