improved accuracy

[Alexander-Barth]

The following changes where made:

1. double the convolutional layers before down-sampling as in the classical VGG networks (https://viso.ai/deep-learning/vgg-very-deep-convolutional-networks/ )
2. use selu as a good compromise between speed and accuracy (https://arxiv.org/pdf/2104.02523.pdf ) avoiding too many untrainable "dead" neurons
3. Use InstanceNorm as first layer to normalize the input
4. Change the saving function to include the state of the normalization layer (see warning of https://github.com/FluxML/Flux.jl/blob/79971741ed8454cdf6a66515799a0c4b864f564a/docs/src/saving.md )
5. increase epochs to 100

I get now about 77% accuracy in my tests.

[rafaqz]

Thanks!!