Can Self-Supervised Representation Learning Methods Withstand Distribution Shifts and Corruptions?