Defining Benchmarks for Continual Few-Shot Learning