Fix PPO's compute_reward() (#996)

[CZH-THU]

If mask is action_mask[:,start:] = attention_mask[:, 1:][:,start:] and ends=start + action_mask[:, start:].sum(1)+1 then rewards[:,ends[j]-1] represent the last non-padding token's reward by predicting padding token which action_mask is 0,The reward_score should give the penultimate non-padding token to reward it's action —— predicting the last non-padding token.