Differentially Private Learning with Per-Sample Adaptive Clipping