Leveraging Locality and Robustness to Achieve Massively Scalable Gaussian Process Regression