TODO

think about using LoopVectorization and KernelAbstractions as (optional?) backends
make k(a) create a broadcastable object (define broadcast style, use windowloop in copy!(bc) and define appropriate axes)
nicer (but type-instable) interface for extensions
abstract/strided array interface for windows (blocked by julia issue)
think about more specific kernel types and composability

done

syntactic sugar for determining kernel size through index access: @kernel(w -> w[1] - w[0]) == Kernel{(0:1,)}(w -> w[1] - w[0])

~~introduce elementary kernel shift(a, CartesianIndex(1, 1)), @shift a[1,1]~~