Integral Signatures of Activation Functions: A 9-Dimensional Taxonomy and Stability Theory for Deep Learning