Failure Modes for Deep Learning-Based Online Mapping: How to Measure and Address Them