Implicit Hypersurface Approximation Capacity in Deep ReLU Networks