RLBench distinguish between train_env and eval_env

Pyrep can only run one instance per process. Therefore we currently have train_env == eval_env for RLBench environments. This has already caused problems and it would be nice if we could distinguish between training and evaluation environment.

Therefore we'd like to

• change the RLBench wrapper so that it is self-aware of its role (train/test) - maybe it is even better to assign a number to each new "instance" of an environment
• remove if episode_timesteps >= self.max_episode_length: break from hac.py, as it is just a quickfix for this problem.