Benchmarking Training Paradigms, Dataset Composition, and Model Scaling for Child ASR in ESPnet