optimise mosaic

[rafaqz]

With all the benchmarking and work on other methods I just want to write this down for the worst offender in the package currently.

mosaic is terrible.

We need to optimise first and last to just write each layer in the right order and use dims2indices to get the indices for each axis rather than working one pixel at a time.

Other methods like sum or mean we can optimise by first calculating the destination indices for each layer as Vector{Int} (with possible duplicate values) for each axis, then going line by line and doing the reduction on iterators of the intersecting layers. This will be really fast. We can also do it over DiskArrays.eachchunk of the destination array.