Allow aggregators to be injected at configuration time

[mcgibbon]

We have a few issues arising from all experiments using the same metrics/aggregators:

• our experiments have a lot of metrics
• it's hard to make changes needed for specific model/experiment types
• it's unclear how an external user would add custom metrics

I'm thinking we should we add a way to "inject" aggregators at configuration time, by specifying the module the class is located in and the configuration arguments to initialize it, with dataset_info automatically passed at init. Then to make metrics specific to an experiment or for an external user, all that's needed is to create the aggregator class (subclassing from an ABC) in an installed package (doesn't need to be ours) and add it to the config.