A Non-isotropic Probabilistic Take on Proxy-based Deep Metric Learning