benchmark different downscaling schemes

[d-v-b]

• compare da.coarsen vs da.map_blocks vs xarray.DataArray.map_blocks
• compare "chained / recursive" downscaling vs "dispersive" downscaling (the current method). Memory usage should be much better for the former approach, but it imposes a more restrictive dependency graph on computations. And it's not clear how chained downscaling can work for nonlinear aggregation methods (e.g., mode) without augmenting the return type (e.g., return an array + counts instead of just an array)