Sparse and bucketized feature weights in GenericLogisticRegressionProto have not been implemented, and we won't need to now that we have a better solution using full_names and preprocessing. We should clean up the code and the proto from these.