BudgetMix RL: What Multi-Turn Reinforcement Learning Really Improves in LLM Agents