From Reasoning to Agentic: Credit Assignment in Reinforcement Learning for Large Language Models