Characterizing Generalization under Out-Of-Distribution Shifts in Deep Metric Learning