Add coarsen.construct

[dcherian]

Like Rolling.construct this would create a new reshaped array with extra dimensions that would usually be reduced over. It would be a light wrapper over Variable._coarsen_reshape

It make #5119, and #2419 really easy without having to go through a MultiIndex+unstack operation which is somewhat arcane.

ds.coarsen(time=12).construct(dim="year").rename({"time": "month"})

[shoyer]

Yes, we should totally do this!

This is a very common operation in machine learning workflows, e.g., in TensorFlow it's tf.image.extract_patches.

[dcherian]

ds.coarsen(time=12).construct(dim="year").rename({"time": "month"})

We could incorporate the renaming like this. Thoughts?

ds.coarsen(time=12).construct(dim={"time": ("month", "year")})