xref pangeo-data#29

This PR moves the dask specific scheduling logic into a separate `dask.py`
file, as a first step for adding support for alternative schedulers. (I'm
particularly interested in supporting Apache Beam.)

The existing tests pass (with minor modifications), but the documentation still
needs updating.

Notes:

- I put `staged_copy` into a single function, but perhaps there are other
  generic methods (`execute`?) that would justify using a class?
- `Rechunked` no longer inherits from `dask.delayed.Delayed`, and no longer has
  any dask specific logic at all. I think this is important for generic
  scheduler support, but it does means make it a little less reusable in larger
  pipelines. `_delayed` is currently a private attribute, but we should
  probably expose the scheduler equivalent of "delayed" objects in some way.
  I guess this is a use-case for class-based interface from the previous
  bullet.
- `Rechunked` now always contains zarr arrays/groups rather than dask arrays.
  This makes the repr a little less informative, e.g., it no longer shows
  chunk size. This should probably be fixed before merging.
- Will "two stage" copying always suffice? The interface I wrote for
  `staged_copy` supports any number of stages (in theory). That might be useful
  in the future, or it might be unnecessary complexity.
- To verify that adding a new scheduler is not too painful, I should probably
  write at least a second example. I'll start with a naive "reference"
  scheduler in pure Python (this could go in the docs) and think about adding
  a Beam implementation as well. Beam is perhaps a nice example because it's
  execution models is so different from dask (based on higher level transforms
  like "map" rather than individual tasks).