Learning Distributions Generated by Single-Layer ReLU Networks in the Presence of Arbitrary Outliers