improve example in README

[marcelluethi]

The example in the README file is not great. This is an attempt to create an example that better illustrates what dimwit can do.

[benikm91]

I like the idea behind this PR; our current normalize example is weak because it shows very little of DimWit. However, while it's an improvement, I don't like the new example. First, it doesn't make sense to me: A single layer with a tanh activation and minimizing the MSE; doesn't the model then just learn to predict 0? Where is this model from? The model also adds explicit Params, which could confuse deep learning folks. We also have a similar example (minimal MLP) in deepwit: https://github.com/dimwit-dev/deepwit

Minor: x.vmap(Axis[Batch])(layer(p)).pow(Tensor0(2.0f)).meancan be simplified to ` x.vmap(Axis[Batch])(layer(p)).pow(2.0).mean?

I will try to come up with an example I like more in the next few days to have a counterproposal.