Efficiently Computable Safety Bounds for Gaussian Processes in Active Learning