nan length coordinates

Is your feature request related to a problem? Please describe.

When using arrays with a nan shape, I'd like to provide a coordinate specification from a delayed object, which is my responsibility to make sure has the right chunks and length.

Describe the solution you'd like

Below are three examples that I think should work, where I give a coordinate via a dask array, a dask series, and a dask index. Currently, all three examples error out by computing the coordinate length (triggering an unwanted dask computation!), and then indicating that its different than the array's corresponding length, which is nan.

import dask
import dask.array as da
import dask.dataframe as dd
import numpy as np
import pandas as pd
import xarray as xr

def foo():
    return np.arange(4)

arr = da.from_delayed(dask.delayed(foo)(), shape=(np.nan,), dtype=int)

idx = da.from_delayed(dask.delayed(foo)(), shape=(np.nan,), dtype=int)

ddf = dd.from_pandas(pd.DataFrame({'y': np.arange(4)}), npartitions=1)

arr0 = xr.DataArray(arr, coords=[('z', idx)])
arr1 = xr.DataArray(arr, coords=[('z', ddf['y'])])
arr2 = xr.DataArray(arr, coords=[('z', ddf.index)])

Error:

ValueError: conflicting sizes for dimension 'z': length nan on the data but length 4 on coordinate 'z'

Describe alternatives you've considered

After computations to complete add the missing coordinate. This requires carrying around the delayed index with the delayed array.

Start by reproducing the three DataArray construction examples in the issue and tracing how coordinates are validated when the data shape contains nan. The work is done when dask array, series, and index coordinates can be attached without computing their lengths or raising a false size conflict.