Cluster bootstrap for metrics; refactor metric computations into evaluate_preds

[norabelrose]

I realized that our code for computing ROC AUROC, accuracy, and calibrated accuracy were sort of all over the place and there was a decent amount of code duplication. This PR refactors all of that into a single function evaluate_preds which is used for both reporters and logistic regression classifiers, in elicit as well as eval.

Other changes:

1. Confidence intervals now use the cluster bootstrap, resampling entire groups of prompt templates at a time, to take account of the fact that different variants of the same data point are not IID. This leads to significantly larger CIs than those reported in main.
2. Partially in order to account for (1), I've increased the default for max_examples from [750, 250] to [1000, 1000]. There's just too much noise in the data with only 250 clusters.
3. Confidence intervals are now included for accuracy and calibrated accuracy

[AlexTMallen]

LGTM