Linear() fallback has non-trivial activation function, whence is non-linear

```julia
julia> MLJFlux.Linear()
Linear(σ = relu)
```

This is a possible a source of confusion, which I encountered in a real analysis. If we change the fallback to `σ = identity`) do we call this a bug fix (non-breaking) or just tag a breaking release? I'm inclined to do the latter. 

