[Reranker] Set decoding method

[jxmsML]

Patch description

1. Add the set_decoding_method for flexibility.
   Seems like for the SeeKeR models, in order to use the --inference-strategies one must set decoding through:
   agent_clone.opt['inference'] = XXXX for agent_clone in self.dialogue_agent_clones
   instead of self.opt['drm_inference'] = XXXX, since self.opt['drm_inference'] is only used during initialization.
2. Add the batch_reply as the parameter of get_observations_for_reranker for flexibility:
   For the SeeKeR case, the knowledge sentence can only be accessed thru batch_reply[i].knowledge_response. Adding batch_reply to the function get_observations_for_reranker would be useful for make the context + knowledge_ sentence as the observation['full_text']

Testing steps

Other information