[PACER] Adding flexibility + removing padding idx before v2t

[jxmsML]

Patch description

1. return Any reranker class not just RPAReranker for partial only agent
2. check if 'classifier' is in shared, this is for the SeeKeR case, where
   dialogue_agent = model_class(opt, knowledge_agent.share()) but knowledge_agent.share() doesn't have classifier attribute
3. adding def get_target_character(self): that returns the target label the reranker tries to optimize for.
4. removing padding id before applying self.agent_v2t(self.context), otherwise during batch generation one might run into this kind of error KeyError, '__null__' not found in self.dict
5. int(pred['text'] == self.character) - 1 replaced with pred['text_candidates'].index(self.character) in case the predictor is a multi-class.

Testing steps

Other information