cuda4dnn(activations, eltwise, scale_shift): refactor to reduce code duplication

* refactor activations

* refactor eltwise kernels

* move all functors to functors.hpp

* remove bias1 and scale1 kernels