Allow more subclassing of self-chat world

[EricMichaelSmith]

Allow for more subclassing of SelfChatWorld to modify/extend its logic:

• Abstract out the logic applied at the end of every episode, in order to allow subclasses to supply custom logic (in my case, resetting a random turn-index variable per episode)
• Allow for specifying custom logic for when to use seed utterances (in my case, only when the current turn index equals the value of the random turn-index variable specified above)

[stephenroller]

maybe we should wrap this into self.reset?

[EricMichaelSmith]

I've created a World.reset_agents() method - is this what you were thinking? We can't call World.reset() directly here, because there's additional logic there - i.e. World.reset() also resets self.total_parleys, self.total_exs, self.time, etc.

[spencerp]

I have a slight preference toward integrating this logic closer with the above:

self.seed_utterances = self._get_seed_utt_acts(
    self.episode_cnt, self.agents
)

Maybe by putting the self._use_seed_utterances check inside of self._get_seed_utt_acts so there's a single source of truth (if self.seed_utterances)?

[EricMichaelSmith]

Yes, good point - changed

[EricMichaelSmith]

Thanks @stephenroller @spencerp for useful comments! I've tried to address your suggestions as best as possible, and since you've both approved the PR, I'm merging it in now - but if you feel like I didn't address your suggestions appropriately, lemme know and I'd be happy to make changes in another PR