Provable Traffic Rule Compliance in Safe Reinforcement Learning on the Open Sea