Continuous vs. dynamic batching for AI inference