option to track avg logit per token type

[IanMagnusson]

This PR is ported from the old LLM repo at allenai/OLMo#334.

This PR makes use of new features in Catwalk's perplexity evaluations in allenai/catwalk#155 that report avg logits for tokens. It makes WriteOutputsAsRowsMultipleMetrics and PredictAndCalculateMetricsStep capable of surfacing these "token_count_avg_logits_by_domain" results within the new "extra_output" field.

To use this one must have the following task kwargs:

task_kwargs: {
    keep_all_instance_fields_except: ["text", "tokens"],
    detailed_output: true
}

You need to use keep_all_instance_fields_except because this depends on a lot of instance level information. And the detailed_output flag indicates that the aggregated results of this instance level data should be surfaced.