About the reward computing logic

[zheangh]

I notice some detailed designs in the _compute_score_with_rules in envs/vision.py or envs/search.py. For instance, when the answer is not extracted, an additional -format_score penalty is imposed, although total_format_score already contains such a punishment. Also, total_format_score is multiplied by 1/2.
I wonder whether the logic in _compute_score_with_rules is based on any reference, or entirely designed by you. Thanks in advance!