New .broadcast_like() incorrect

[DavidMertz]

Unfortunately, more robust tests show there is some logic problem in the @dcherian solution to #2885. I was merging my work in #3127 with his more-principled refactoring in #3086, and my tests showed an error in the refactoring.

[DavidMertz]

In https://github.com/DavidMertz/xarray/tree/feature/2885-broadcast_like-merge-branch I have created a temporary method .broadcast_like_naive() and have created a test against the naive implementation that is identical to the test against the refactored code.

We actually get the failure in the call to .broadcast_like() itself, i.e. not in the assertion about the result:

___________________________________________ TestDataArray.test_broadcast_like ____________________________________________

self = <xarray.tests.test_dataarray.TestDataArray object at 0x7f2287e3a828>

    def test_broadcast_like(self):
        arr1 = DataArray(np.ones((2, 3)), dims=['x', 'y'],
                         coords={'x': ['a', 'b'], 'y': ['a', 'b', 'c']})
        arr2 = DataArray(np.ones((3, 2)), dims=['x', 'y'],
                         coords={'x': ['a', 'b', 'c'], 'y': ['a', 'b']})
        orig1, orig2 = broadcast(arr1, arr2)
>       new1 = arr2.broadcast_like(arr1)

xarray/tests/test_dataarray.py:1276: 
[...]
E                   ValueError: conflicting sizes for dimension 'y': length 2 on the data but length 3 on coordinate 'y'

[DavidMertz]

With the following method instead, the identical test succeeds (modulo using the temporary method for tests, of course):

    def broadcast_like_naive(self,
                       other: Union['DataArray', Dataset],
                       exclude=None) -> 'DataArray':
        return broadcast(other, self)[1]