[Crowdsourcing] edits for OSS personal knowledge project

[jxmsML]

Patch description

• Add ContextGenerator based on the task_name = 'msc'
• Flexibility to build ContextGenerator for any given task:
  • Modify conversation_start_mode to be either 'hi', or a 'task name' (e.g. 'blended_skill_talk', or 'msc'
  • get_context_generator in parlai/crowdsourcing/tasks/model_chat/utils.py will load ContextGenerator from the <task_name>/agents.py specified by conversation_start_mode
• Add world_module as an additional parameter for run_task in parlai/crowdsourcing/tasks/model_chat/run.py
• In parlai/crowdsourcing/tasks/model_chat/worlds.py, display fancy bot name with self.bot.agent_id rather than the default agent name from the 'id' field (mostly TransformerGenerator)
• Move "validate problem buckets" from AbstractTurnAnnotationResultsCompiler to ModelChatResultsCompiler init method.

if 'none_all_good' not in self.problem_buckets:
        # The code relies on a catchall "none" category if the user selects no other
         # annotation bucket
        raise ValueError(
       'There must be a "none_all_good" category in self.problem_buckets!'
         )

Testing steps
CI

Other information

[jxmsML]

move out of parlai/crowdsourcing/utils/analysis.py here as it could be that "none_all_good" isn't necessary in problem_buckets.

[EricMichaelSmith]

hmm, okay, on closer look, I'm fine with moving this check into the subclasses, because the logic that relies upon "none_all_good" being present is itself in the subclasses

[jxmsML]

adding world_module here so it doesn't have to call the model_chat.worlds by default but any world_module as long as it's consistent with cfg.mephisto.blueprint.world_file.

[EricMichaelSmith]

Yeah, good to have - thanks for generalizing run_task, etc.! I have a few nitpicky questions, and in particular, before approving, it might be good to decide where the 'none_all_good' check should go