id,node_id,number,title,user,state,locked,assignee,milestone,comments,created_at,updated_at,closed_at,author_association,active_lock_reason,draft,pull_request,body,reactions,performed_via_github_app,state_reason,repo,type
775875024,MDU6SXNzdWU3NzU4NzUwMjQ=,4739,Slow initilization of dataset.interp,14371165,closed,0,,,2,2020-12-29T12:46:05Z,2021-05-05T12:26:01Z,2021-05-05T12:26:01Z,MEMBER,,,,"<!-- Please include a self-contained copy-pastable example that generates the issue if possible.

Please be concise with code posted. See guidelines below on how to provide a good bug report:

- Craft Minimal Bug Reports: http://matthewrocklin.com/blog/work/2018/02/28/minimal-bug-reports
- Minimal Complete Verifiable Examples: https://stackoverflow.com/help/mcve

Bug reports that follow these guidelines are easier to diagnose, and so are often handled much more quickly.
-->

**What happened**:
When interpolating a dataset with >2000 dask variables a lot of time is spent in `da.unifying_chunks` because `da.unifying_chunks` forces all variables and **coordinates** to a dask array. 
xarray on the other hand forces coordinates to pd.Index even if the coordinates was dask.array when the dataset was first created.

**What you expected to happen**:
If the coords of the dataset was initialized as dask arrays they should stay lazy.

**Minimal Complete Verifiable Example**:

```python
import xarray as xr
import numpy as np
import dask.array as da

a = np.arange(0, 2000)
b = np.core.defchararray.add(""long_variable_name"", a.astype(str))
coords = dict(time=da.array([0, 1]))
data_vars = dict()
for v in b:
    data_vars[v] = xr.DataArray(
        name=v,
        data=da.array([3, 4]),
        dims=[""time""],
        coords=coords
    )
ds0 = xr.Dataset(data_vars)
ds0 = ds0.interp(
    time=da.array([0, 0.5, 1]),
    assume_sorted=True,
    kwargs=dict(fill_value=None),
)
```

**Anything else we need to know?**:
Some thoughts:
* Why can't coordinates be lazy? 
  * Can we use dask.dataframe.Index instead of pd.Index when creating IndexVariables?
* There's no time saved converting to dask arrays in `missing.interp_func`. But some time could be saved if we could convert them to dask arrays in `xr.Dataset.interp` before the variable loop starts. 
* Can we still store the dask array in IndexVariable and use a to_dask_array()-method to quickly get it? 
  * Initializing the dataarrays will still be slow though since it still has to force the dask array to pd.Index.


**Environment**:

<details><summary>Output of <tt>xr.show_versions()</tt></summary>

<!-- Paste the output here xr.show_versions() here -->
xr.show_versions()

INSTALLED VERSIONS
------------------
commit: None
python: 3.8.5 (default, Sep  3 2020, 21:29:08) [MSC v.1916 64 bit (AMD64)]
python-bits: 64
OS: Windows
OS-release: 10
libhdf5: 1.10.4
libnetcdf: None

xarray: 0.16.2
pandas: 1.1.5
numpy: 1.17.5
scipy: 1.4.1
netCDF4: None
pydap: None
h5netcdf: None
h5py: 2.10.0
Nio: None
zarr: None
cftime: None
nc_time_axis: None
PseudoNetCDF: None
rasterio: None
cfgrib: None
iris: None
bottleneck: 1.3.2
dask: 2020.12.0
distributed: 2020.12.0
matplotlib: 3.3.2
cartopy: None
seaborn: 0.11.1
numbagg: None
pint: None
setuptools: 51.0.0.post20201207
pip: 20.3.3
conda: 4.9.2
pytest: 6.2.1
IPython: 7.19.0
sphinx: 3.4.0

</details>
","{""url"": ""https://api.github.com/repos/pydata/xarray/issues/4739/reactions"", ""total_count"": 0, ""+1"": 0, ""-1"": 0, ""laugh"": 0, ""hooray"": 0, ""confused"": 0, ""heart"": 0, ""rocket"": 0, ""eyes"": 0}",,completed,13221727,issue