Hybrid activation functions for deep neural networks: S3 and S4 -- a novel approach to gradient flow optimization