Invalid start states used for report method and can "carry" be eliminated in train/report methods

[sai-prasanna]

1. Invalid start state in report
   In the recent changes, we use the context from the replay buffer when self.config.replay_context > 0. But for report method, currently, the code always passes the start state zeros context as carry. Maybe we mirror what's done in train and use the replay context if it is set?

2. I don't see the point of carry for train and report methods. I understand that we might need it for online training, where we are observing one observation (or batch of single timestep observations) at each train step and returning the encoded posterior after train step. But I think this is purely conceptual and does not serve any purpose right now?