[BUG] The advantage computing is wrong, even naive advantage should be R(t)-V(t), R is Return, not reward

[schinger]

https://github.com/hpcaitech/ColossalAI/blob/285fe7ba7183f16442b570ac3e0f2de0e567d009/applications/Chat/coati/experience_maker/naive.py#L52C35-L52C35

[CWHer]

This is kind of duplicated. Maybe we can discuss this in another open issue. :)

Please check this #4475 (comment)