Learning to Refine: An Agentic RL Approach for Iterative SPARQL Query Construction