Custom evaluation metrics

[ywkim]

This PR lets custom evaluation metrics to be defined at problem level.

It would be nice if we could add custom metrics to problems. (See also #822) I would like to propose a new method to the Problem that returns metric functions.

def eval_metric_fns(self, model_hparams):
  metric_names = self.eval_metrics()
  return {
      metric_name: metrics.METRICS_FNS[metric_name]
      for metric_name in metric_names
  }

[afrozenator]

Thanks a lot for doing this @ywkim ! This is a very good idea and we had multiple folks requesting this internally as well -- so thanks a lot again for doing this!