Whether a DataArray is copied when inserted into a Dataset depends on whether coordinates match exactly

Consider

import numpy as np
from xray import *

ds = Dataset({"a": DataArray(np.zeros((3, 4)))})
ds["b"] = b = DataArray(np.zeros((3, 4)))
b[0, 0] = 1
print(ds["b"][0, 0]) # ==> prints 1

ds = Dataset({"a": DataArray(np.zeros((3, 4)))})
ds["b"] = b = DataArray(np.zeros((3, 3)))  # !!! we implicitly fill the last column with nans.
b[0, 0] = 1
print(ds["b"][0, 0]) # ==> prints 0

In the first case, the dataset was modified when the dataarray was modified, but not in the second case.

Reproduce both examples from the issue and inspect the Dataset insertion and coordinate-alignment paths involved in assigning a DataArray. No file or test is named; done should mean the copy behavior for matching and nonmatching coordinates is consistent and covered by a regression test.