Adaptive Batch Size Schedules for Distributed Training of Language Models with Data and Model Parallelism