This is a follow-up to generalize the changes from pydata#1024:

- Caching and copy-on-write behavior has been moved to separate array classes
  that are explicitly used in `open_dataset` to wrap arrays loaded from disk (if
  `cache=True`).
- Dask specific logic has been removed from the caching/loading logic on
  `xarray.Variable`.
- Pickle no longer caches automatically under any circumstances.

Still needs tests for the `cache` argument to `open_dataset`, but everything
else seems to be working.