Symmetrisation should be a LinearLayer

[cortner]

julia> l_basis.layers.BB.layers.layer_1
WrappedFunction(#82)

... and not a wrapped Function. In general, please avoid use of WrappedFunction as much as possible. It makes for unreadable code and it will be hellish when we start to serialize and deserialize models.

[cortner]

It seems this wrapped function

julia> l_basis.layers.BB
Parallel(
    layer_1 = WrappedFunction(#97),
)         # Total: 0 parameters,
          #        plus 0 states.

is currently preventing reverse over reverse differentiation. It would be good to replace it. I'll have to use some hack in the meantime.