Improving Moleculenet's metrics

MoleculeNet has some limitations in its original choice of metrics. Here are a couple issues:

- Currently MoleculeNet only tests RMS for its regression tasks. For some tasks, RMS isn't a good measure, and something like R^2 would be more suited for some tasks. (H/T @PatWalters for bringing this up).
- For a number of tasks involving molecular classification, auPRC is a better metric than ROC-AUC. Having auPRC added to the classification sections would be very useful.