Reinforcing Multi-Turn Reasoning in LLM Agents via Turn-Level Credit Assignment