Pruning Neural Networks via Coresets and Convex Geometry: Towards No Assumptions