Probabilistic Safeguard for Reinforcement Learning Using Safety Index Guided Gaussian Process Models