implement normalize_token

[dcherian]

See #3276 (comment)

[dcherian]

How should this be implemented?

[crusaderky]

https://docs.dask.org/en/latest/custom-collections.html#implementing-deterministic-hashing

@normalize_token.register(Dataset)
def tokenize_dataset(ds):
    return Dataset, ds._variables, ds._coord_names, ds._attrs

@normalize_token.register(DataArray)
def tokenize_dataarray(da):
    return DataArray, ds._variable, ds._coords, ds._name

# Note: the @singledispatch for IndexVariable must be defined before the one for Variable
@normalize_token.register(IndexVariable)
def tokenize_indexvariable(v):
    # Don't waste time converting pd.Index to np.ndarray
    return IndexVariable, v._dims, v._data.array, v._attrs

@normalize_token.register(Variable)
def tokenize_variable(v):
    # Note: it's v.data, not v._data, in order to cope with the 
    # wrappers around NetCDF and the like
    return Variable, v._dims, v.data, v._attrs

You'll need to write a dummy normalize_token for when dask is not installed.

Unit tests:

• running tokenize() twice on the same object returns the same result
• changing the content of a data_var (or the variable, for DataArray) changes the output
• changing the content of a coord changes the output
• changing attrs, name, or dimension names change the output
• whether a variable is a data_var or a coord changes the output
• dask arrays aren't computed
• non-numpy, non-dask NEP18 data is not converted to numpy
• works with xarray's fancy wrappers around NetCDF and the like