to_zarr() failing after concatenating netcdfs with different time indexes

What happened:
After concatenating two NetCDF DataSets with different cftime.DateTimeNoLeap coordinates, attempting to write to a Zarr store with ds.to_zarr() fails with an OutOfBoundsDatetime exception.

What you expected to happen:
I expect to_zarr() to execute successfully.

Minimal Complete Verifiable Example:

import xarray as xr
import cftime
import pandas as pd

# open a generic CESM dataset containing a time_bnds variable
url = 'http://adss.apcc21.org/opendap/CMIP5DB/cmip5_daily_BT/pr_day_CESM1-BGC_rcp85_r1i1p1_20760101-21001231.nc'
ds = xr.open_dataset(url)

# create two new DataSets with different, overlapping time indexes.
ds2 = ds.sel(time=slice(None, cftime.DatetimeNoLeap(2076, 3, 1, 1, 0, 0, 0)))
ds3 = ds.sel(time=slice(None, cftime.DatetimeNoLeap(2076, 2, 1, 1, 0, 0, 0)))

# concatenate the two DataSets, using the default fillvalue
ds4 = xr.concat([ds2, ds3], dim=pd.Index(['ds2','ds3'], name='ds'))

# fails with OutOfBoundsDatetime exception
zs = ds4.to_zarr('/tmp/my_zarr.zarr')

Anything else we need to know?:
I believe the problem is related to the implicit NaN fillvalue used in concatenating the time_bnds variable. I arrived at the code while trying to produce a minimal example of a similar error, where a DataSet concatenation would fail with a SerializationError if I tried to concatenate multiple datasets containing time_bnds variables. In the above example and in my earlier troubleshooting in production code, removing time_bnds with ds = ds.drop('time_bnds') made the to_zarr() command work.

This is a common use case since time_bnds indexes are generated by CESM climate model output.

Environment:

Start by running the minimal example with xr.concat and ds4.to_zarr(), keeping the time_bnds variable, and confirm the OutOfBoundsDatetime failure. Trace the serialization path used by to_zarr for the concatenated time_bnds data; done means the example writes successfully to Zarr without requiring ds.drop('time_bnds').