DIVEBATCH: Accelerating Model Training Through Gradient-Diversity Aware Batch Size Adaptation