Delayed-like functionality

[tomwhite]

There are some cases where it would be useful to have a way of setting up a parallel map over inputs, then calling compute on it later - like Dask Delayed.

In particular this would be needed to support xarray's open_mfdataset (see pydata/xarray#7811), or creating Kerchunk files in parallel.

The main question is what this looks like and where it belongs.

[tomwhite]

Here's a basic proof of concept: 5f6e38e