Adapting Open-Source Large Language Models for Cost-Effective, Expert-Level Clinical Note Generation with On-Policy Reinforcement Learning