Support decision tasks by providing reward wrapper for Gym-like RL environment

[MA-Wenhui]

What does this PR do?

Support decision tasks, by providing an Environment reward wrapper, such as a Gymnasium-like RL environment.

When training ppo model for decision task, other than using a reward model, one could use his own environment system code to generate rewards by specifying following params in yaml:

reward_model: your_env_package.reward_model_wrapper.SampleRewardModel
reward_model_type: env

The SampleRewardModel is within user's environment package, it does:

1. parses the Query and Response
2. executes action, interacts with environment
3. returns reward

The SampleRewardModel example is :

class RewardModel:
    def __init__(self):
        self.env = Environment()

    def __call__(self, query_texts, response_texts):
        # do environment actions and get reward
        return self.get_rewards(query_texts, response_texts)

Before submitting

• [ yes ] Did you read the contributor guideline?
• [ yes ] Did you write any new necessary tests?
  A new ppo example sample yaml