Use model name and dataset to organize reporters in elicit

[norabelrose]

We're still using memorable names to disambiguate different elicit runs with different hyperparameters. But these are now nested inside a much more organized structure like this:

~/elk-reporters/bigscience/bloomz-560m/sst2/objective-yalow

for a reporter trained on bigscience/bloomz-560m on sst2. This makes it much easier to track down the reporter you're looking for, especially when generating reporters over tons of different (model, dataset) combinations.