PLRV-O: Advancing Differentially Private Deep Learning via Privacy Loss Random Variable Optimization | Synapse