Rewrite DataTree.to_netcdf and support netCDF4 in-memory

[shoyer]

This PR includes a handful of significant changes:

1. It refactors the internal structure of DataTree.to_netcdf() and DataTree.to_zarr() to use lower level interfaces, rather than calling Dataset methods. This allows for properly supporting compute=False (and likely various other improvements).
2. Reading and writing in-memory data with netCDF4-python is now supported, including DataTree.
3. I've added a new user-facing load_datatree function, for consistentcy with load_dataset and load_dataarray.
4. The engine argument in DataTree.to_netcdf() is now set consistently with Dataset.to_netcdf(), preferring netcdf4 to h5netcdf.
5. Calling Dataset.to_netcdf() without a target now always returns a memoryview object, including in the case where engine='scipy' is used (which currently returns bytes). This is a breaking change, rather than merely issuing a warning as is done in Support for DataTree.to_netcdf to write to a file-like object or bytes #10571. I believe it probably makes sense to do as a this breaking change because (1) it offers significant performance benefits, (2) the default behavior without specifying an engine will already change (because netcdf4 is preferred to the scipy backend) and (3) restoring previous behavior is easy (by wrapping the memoryview with bytes()).

• Tests added
• User visible changes (including notable bug fixes) are documented in whats-new.rst

[shoyer]

• It refactors the internal structure of DataTree.to_netcdf() and DataTree.to_zarr() to use lower level interfaces, rather than calling Dataset methods. This allows for properly supporting compute=False (and likely various other improvements).

I am thinking I might try to split this into a separate PR, because it's unrelated to the netCDF4 in-memory changes.