segmentation fault when numpy_npz data_reader used with validation_percent != 0.0

[whitesides1]

Trying to use the numpy_npz reader with validation_percent != 0.0 leads to a segmentation fault in proto_common.cpp. A "split_reader" pointer is created and then never set and no check is made that the "name" matches one of the supported readers for splitting the data. A fall-through check should be made on name to make sure that split_reader is set to a valid reader. It would also be nice if numpy_npz reader could be made to support the validation split.