Scalable Partial Explainability in Neural Networks via Flexible Activation Functions