Fix groupby binary ops when grouped array is subset relative to other

[dcherian]

• Closes More groupby indexing problems #7797
• Tests added
• User visible changes (including notable bug fixes) are documented in whats-new.rst

cc @slevang

[headtr1ck]

I think the broken test was actually wrong before, now the actual result looks more like what I would expect.

[slevang]

Thanks for the quick fix! Not sure about the bitshift test but I'm assuming @headtr1ck is right.

[abrammer]

Apologies, that's my bad. Looks like I introduced a broken test and didn't manually double check the results coming back. The right shift test should have been:

    right_expected = Dataset(
        {
            "x": ("index", [0, 0, 2, 2]),
            "y": ("index", [-1, -1, -2, -2]),
            "level": ("index", [0, 0, 4, 4]),
            "index": [0, 1, 2, 3],
        }
    )

    right_actual = (left_expected.groupby("level") >> shift).reset_coords(names="level")
    assert_equal(right_expected, right_actual)

I haven't paid attention to this issue, but doing the groupby manually didn't have the bug fwiw.

Probably overkill test that only fails at the last assert before this fix

def test_groupby_math_bitshift() -> None:
    # create new dataset of int's only
    ds = Dataset(
        {
            "x": ("index", np.ones(4, dtype=int)),
            "y": ("index", np.ones(4, dtype=int) * -1),
            "level": ("index", [1, 1, 2, 2]),
            "index": [0, 1, 2, 3],
        }
    )
    shift = DataArray([1, 2, 1], [("level", [1, 2, 8])])

    left_expected = Dataset(
        {
            "x": ("index", [2, 2, 4, 4]),
            "y": ("index", [-2, -2, -4, -4]),
            "level": ("index", [2, 2, 8, 8]),
            "index": [0, 1, 2, 3],
        }
    )

    left_manual = []
    for lev, group in ds.groupby("level"):
        shifter = shift.sel(level=lev)
        left_manual.append(group << shifter)
    left_actual = xr.concat(left_manual, dim="index").reset_coords(names="level")
    assert_equal(left_expected, left_actual)

    left_actual = (ds.groupby("level") << shift).reset_coords(names="level")
    assert_equal(left_expected, left_actual)

    right_expected = Dataset(
        {
            "x": ("index", [0, 0, 2, 2]),
            "y": ("index", [-1, -1, -2, -2]),
            "level": ("index", [0, 0, 4, 4]),
            "index": [0, 1, 2, 3],
        }
    )
    right_manual = []
    for lev, group in left_expected.groupby("level"):
        shifter = shift.sel(level=lev)
        right_manual.append(group >> shifter)
    right_actual = xr.concat(right_manual, dim="index").reset_coords(names="level")
    assert_equal(right_expected, right_actual)

    right_actual = (left_expected.groupby("level") >> shift).reset_coords(names="level")
    assert_equal(right_expected, right_actual)

[slevang]

Would it be possible to run another bug fix release with this incorporated? Or I guess we're already on to 2023.5.0 according to the date.